From mboxrd@z Thu Jan 1 00:00:00 1970 Path: news.gmane.org!not-for-mail From: Stefan Israelsson Tampe Newsgroups: gmane.lisp.guile.devel Subject: redo-safe-variables and redo-safe-parameters Date: Thu, 04 Apr 2013 23:13:08 +0200 Message-ID: <12760509.JQCs2Vm7n7@warperdoze> NNTP-Posting-Host: plane.gmane.org Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7Bit X-Trace: ger.gmane.org 1365110016 21614 80.91.229.3 (4 Apr 2013 21:13:36 GMT) X-Complaints-To: usenet@ger.gmane.org NNTP-Posting-Date: Thu, 4 Apr 2013 21:13:36 +0000 (UTC) To: guile-devel@gnu.org, noah.b.lavine@gmail.com Original-X-From: guile-devel-bounces+guile-devel=m.gmane.org@gnu.org Thu Apr 04 23:14:00 2013 Return-path: Envelope-to: guile-devel@m.gmane.org Original-Received: from lists.gnu.org ([208.118.235.17]) by plane.gmane.org with esmtp (Exim 4.69) (envelope-from ) id 1UNrTp-0008JU-UH for guile-devel@m.gmane.org; Thu, 04 Apr 2013 23:13:58 +0200 Original-Received: from localhost ([::1]:44502 helo=lists.gnu.org) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1UNrTQ-0001EV-Uu for guile-devel@m.gmane.org; Thu, 04 Apr 2013 17:13:32 -0400 Original-Received: from eggs.gnu.org ([208.118.235.92]:49717) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1UNrTG-0001Dn-CG for guile-devel@gnu.org; Thu, 04 Apr 2013 17:13:28 -0400 Original-Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1UNrTA-0008HR-0I for guile-devel@gnu.org; Thu, 04 Apr 2013 17:13:22 -0400 Original-Received: from mail-lb0-f169.google.com ([209.85.217.169]:43515) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1UNrT9-0008H7-Fa for guile-devel@gnu.org; Thu, 04 Apr 2013 17:13:15 -0400 Original-Received: by mail-lb0-f169.google.com with SMTP id p11so3187336lbi.0 for ; Thu, 04 Apr 2013 14:13:14 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=x-received:from:to:subject:date:message-id:user-agent:mime-version :content-transfer-encoding:content-type; bh=M7AhESqaiiTuwUPO+QkkD+b+xKbNs4VnI42O848P0+0=; b=h9BpMai36QBWuOJxlnP9+yS5Cc0d7PhdCmxTxj7M0Em+3zTVs/HJ1xH5NKCgZnWNle wnRplxVKG29nouxM9A4LLdj0tPSBen7q+pnyZPxPVXJbHFJ1nG4rebaLxNBxK2P0X1+M jTQqd0pAjycX8+IYVMog1i5eo7PK2IgNc/BDyTv1RVJ8lwSlyxsqHx4q3HPYzIHxxwGS BHeAs4pXqUcUMyRQkX0XfO6TWXNJ0wdY6fXONopMRmkJU07Qnt1LuY2C7Ur6ocaOMf4W mMkiEd7GuhoN8Hgoy9dyL9Dneo10uLp/F5LSPql7T86C/7GL5AKW9s5FYpSebWXZ8Am5 jHEQ== X-Received: by 10.152.134.164 with SMTP id pl4mr4424328lab.54.1365109994329; Thu, 04 Apr 2013 14:13:14 -0700 (PDT) Original-Received: from warperdoze.localnet (1-1-1-39a.veo.vs.bostream.se. [82.182.254.46]) by mx.google.com with ESMTPS id hk10sm4659486lab.4.2013.04.04.14.13.12 (version=TLSv1.1 cipher=ECDHE-RSA-RC4-SHA bits=128/128); Thu, 04 Apr 2013 14:13:13 -0700 (PDT) User-Agent: KMail/4.9.5 (Linux/3.5.0-26-generic; KDE/4.9.5; x86_64; ; ) X-detected-operating-system: by eggs.gnu.org: GNU/Linux 3.x [fuzzy] X-Received-From: 209.85.217.169 X-BeenThere: guile-devel@gnu.org X-Mailman-Version: 2.1.14 Precedence: list List-Id: "Developers list for Guile, the GNU extensibility library" List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: guile-devel-bounces+guile-devel=m.gmane.org@gnu.org Original-Sender: guile-devel-bounces+guile-devel=m.gmane.org@gnu.org Xref: news.gmane.org gmane.lisp.guile.devel:16142 Archived-At: Hi, Ok, here is a new version. i tried to refine it and meet your arguments with good actions. We are not there yet. But after actually studying these things in guile-log I'm possitive that it starts to shape up to a sane proposal. WDYT? Dear editors of srfi, Summary: ======== Reinstating continuation in order to capture a state and later restart from the same state is a concept that at least in proof assistants and similar software with user interaction is important. But also direct algorithms in logic programming can take advantage of this concept. Unfourtunately scheme is partly broken when trying to implement these constructs effectively with little effort and hence it could be an idea to try improve on the issue. This srfi describes semantics that can help improve the current state without sacrificing code clearity and the elegance of scheme. We will refere to this concept as redo semantics and hence name the new variable types as redo safe varibles and a new parameter type redo safe parameters. Also we should distinguish between differnt kinds of undo. Generally the supported ideom should be that of having a base function that call utility functions that may end nonlocally with a continuation argument and allow to reinstate that continuation with no change of internal state. Contrast this with executing a utility function that get's a continuation as a parameter and when it executes that function moves only backwards to the continuation point. Redo safe here will mean installing a continuation at backtracking and crosses e.g. dynamic wind barriers. Authors background: I'm an active developer of scheme centered around the guile scheme ecosystem and the developer of guile-log, a logic programming system on top of guile. A short background on the issue: The background has a plurality of origins that converged into one concept. 1. Current scheme is a beautiful language but has a drawback. If you set! a variable that variable will be boxed and any uses of redo semantic becomes broken. This is not acute in scheme because we try to avoid set! and that is well. But there are cases where you would like to use set!. One example being utilizing a simple generator e.g. (next! n) -> (begin (set! n (+ n 1)) n) This can many times be optimized efficiently and with a good semantic in a looping macro for example but it will not be scheme. To achieve that without using set! seams to be difficult. Also the resulting loop! macro will not mix well with redo features on top of the system. 2. Emacs lisp is coded very much in an imperative style and with guile we would like to sell guile scheme as a backend for emacs lisp and emacs as a whole. One of the selling point to do this could be that we could use continuations to seamlessly add undo redo like semantics to already written programs in emacs lisp. The problem is twofold here. i) Typically emacs variable are dynamical variables. Dynamic variables in scheme is modeled on parameters and they posses the property that for each dynamic state S and parameter p the evaluation of p under S is always the last value setted e.g. it is working like a local variable in the S environment. Because each instantiation of a continuation get's it's own dynamic state, the possibility of freezing a state that can be restarted multiple times seams to be there assuming there is no other non local exits. The problem is that a dynamic state can be referenced as a scheme object and used as the current state multiple times as well for the same continuation and hence we must overwrite the initial value at a normal return from the dynamic state and hence loose the ability to store and restore naturally. This property also prevent a possible optimization to store the parameter directly on the stack. ii) Scheme local variables does also show up in the compilation of emacs lisp to guile. The problem with this is that typically lisp programs in emacs is littered with set!-ing of local variables. And the semantic for this in scheme is to box them and again you loose the ability to redo and undo without a hefty reprogramming or write an advanced compiler to scheme - which is typically not a beautiful solution. Again a good concept for enabling seamless and efficient and well designed code that can be undone and redone multiple times is needed. 3. Guile log is a combination of both traditional stack based logic programming concept and the kanren concept and together with a separate continuation implementation targeted for logical variables. To be able to code all of kanrens concept including a stack as well for speed and extra semantics like easy tracing or more generally to have a dynamic-wind like idioms. The ability to redo and undo is really helpful and enables interesting logic programs to be written. So here it was found out that having variables that dynamically can change between behaving like normal variables or during other dynamical context behave like a state preserving variable. The problem is that everything was based on a separate stack implementation instead of reusing the internal ones in guile scheme. In short it would be nice to use less application code and take advantage of fundamental scheme concepts with classy implementation semantics. It is actually possible with current scheme + parameters to implement these concepts, but the code will be slow and risk to be hard to reason about. Rational for lifting this to a srfi request: The concept have been proven useful and also implementing this in the right way do touch on the fundamentals of scheme in such a way that it, from the perspective of the author, needs help from the greater scheme community in order to improve upon the idea leading to a good specification. Also having it as a srfi will mean that redo semantics will have the potential to be better served in a portable manner. Main Caveats: 1. It is possible to implement functionality in r5rs + parameters, but the ideoms runs 10x or more slower then an effective implementation. And result in bloated code. 2. The functionality without a good language support risk of messing up the ability to reason about scheme code. Therefore we want support in scheme to make it possible to introduce this concept in a sane way. Suggested Spec: In order to specify the logic we need: 1. dynamic-wind from r5rs. call/cc from r5rs, and parameters from srfi-39. 2. It is suggested a new variable kind will be called a redo safe variable and a new parameter object called redo safe parameter and the concept will be marked with a ~, like with ! a macro or function that spread this concept should be marked with ~. Else if the semantic follow what is expected by the standard use the usual symbol nomenclature. 3. We will add two parameters that will be bound to a predicates meaning of the variable between a normal and redo safe state. We will introduce them as, redo-variable-wind-guard? redo-parameter-wind-guard? They have the restrictions: the value are evaluated to #t if v is #t and #f when v is #f. 4. There should be a setter and a getter of the predicate functions. (redo-wind-predicate-set! f) (redo-wind-predicate-ref) (redo-wind-parameter-predicate-set! f) (redo-wind-parameter-predicate-ref) There is no semantics connected to the values that the function gives on #f and #t. 5. We need to index a storage with continuation k, id and a symbol s, the continuation and the id is referenceing the storage location. E.g. assuming. (storage-ref k i s) (storage-set! k i s v) where the location will be realeased if k and i is no longer referenced. 6. A redo safe operation is done by storing a current continuation while jumping by calling another continuation keeping the stored continuation as a reference. We will refere to the stored continuation as k. 7. A ordinary variable could be blessed to a redo safe variable by the idiom (with-redo-variables ((s v) ...) code ...), with s ... being variable identifiers, v ... scheme expressions which are evaluated at the evaluation of the form and code ... is the code where s will (possibly) be redo safe in the meaning that if a redo safe operation crosses the with-redo-variables boundary then we will store the value of s ... via a (storage-set! k id 's s) ... operation where id is an object that represent the dynamic identity of the form. If k is instantiated then we will restore s ... via (set! s (storage-ref k id 's)) if (redo-variable-wind-guard?) evaluated with v is not #f. 8. A parameter like scheme object that behaves well with redo and undo will be called a redo safe parameter. 9. A ordinary parameters could be made a redo safe parameters by the idiom (with-redo-parameters ((p u v) ...) code ...), with p ... being evaluated to parameter objects, u ... scheme expressions represented the usual parameter values and v ... again scheme expresions that evaluates to control objects to direct the usage of redo semantics for the parameter dynamically. All these expresions is evaluated at the evaluation of the form e.g. there is an imlicit let, the values is then used in the logic below. code ... is the code where p will be (possibly) redo safe. The semantic is that as with parameterize, but with the addition that if a redo safe operation crosses the with-redo-prameters boundary then we will store the value of s ... via a (storage-set! k id 's (p)) ... operation where id is an object that represent the dynamic identity of the form. If k is instantiated then we will restore s ... via (p (storage-ref k id 's)) in the dynamic context of code ... e.g. after the parametrize part of with-redo-parameters have been done. Again nothing is restored if (redo-parameter-wind-guard?) evaluated on the value of v is not #f. 10) We introduce (set~ a v) for setting redo safe variables and (~ a) to refere to redo safe variables, refering to a redo safe variable in any other way is an error. 11) A redo safe parameter is referenced by (p~ref p) and set by (p~set! p) and refere directly to the object via (~ p). Refering to the object in any other way is an error. We will add to this a specification to support tail call's. 11. Lexically provable tail-call property. Consider the sorce code beeing inlined as much as theoretically possible. Define property S as. A closure object has property S if it includes an object of property S or ~. The return value of a function has property S if any of the arguments, the function included has property S. redo-safe-parameters have property S. In variable bindings the S property carries over to the binding identifier. The value of (relax-s-property c) does not have the S property whatever the property of c. 12. If the with-redo-... construct is in tail call position and a function call is located at a tail call position in code ... then if none of the arguments, the funciton included has the S property, the call should be a tail call. 13. If the a continuation is never referenced outside of the internals during the body code ... for any of the constructs, then a call in tail-call position should be a proper tail-call. Rational and discussion: a) the rational for #t passthrough is to help the important case where we know that we want to store variables for a redo. after optimising redo properties. Putting it to #t and noting that ~ variables are never placed lexically in a lambda, and the variable beeing local together with a proof that every backtracking does not end up in any local visible place we can just use unboxed values and set the values directly on the stack. #f is really just to be consistant and an opertunity for macros to turn off the construct. b) The rational behind specifying that we store and restore at a dynamic-wind position in stead of having the storage global is that it will be impossible to take any short-cut's when storing the state at the creation of a continuation. One will essentially need to store every variable that is in the dynamic scope and have been placed in a closure. This can cause severe computatinal complexity to many algorithm. The main use cases refere mainly to partial-closures which naturally lead to redo safe operations. E.g. the user call's a funciton which ends non locally to a handler where there is user interaction and/or algorithmic descitions and then the continuation is evaluated in order to restart from the old state. It is true though that in some cases it would have been convinient to be able to store and restore the state with the full meaning of the notion. But most of those scenarios could be transformed into a quick jump back and forth in order to store the needed state. c) Redo safe variables behaves well when one delay evaluation e.g. and are erally different than redo safe parameters which main use would be to make sure that parameters can be stored and restored in code that uses them e.g. in emacs. Example 1: (define-syntax-rule (next! i) (begin (set! i (+ i 1)) i)) (define-syntax-rule (next~ i) (begin (set~ i (+ (~ i) 1)) (~ i))) (define-syntax for (syntax-rules (~) ((_ ((~ x) from k) code ...) (let ((x (- k 1))) (with-redo-variables ((x #t)) (let loop () (set~ x (+ (~ x) 1)) code ...)))) ((_ (x from k) code ...) (let ((x (- k 1))) (let loop () (set! x (+ x 1)) code ...))) (for ((~ x) from 0) (f (~ x))) ;; redo safe, behaves as if x is just a local variable that is never ;; set!-ed. The code is well optimized and as fast as if non boxed ;; values are used (for (x from 0) (f (lambda () (+ x (next! x))))) ;; not redo safe, the code works as if set! is used under the hood (for ((~ x) from 0) (f (lambda () (+ x (next~ x))))) ;; Error, user is confused about mixing ~ context with non ~ context (for ((~ x) from 0) (f (lambda () (+ (~ x) (next~ x))))) ;; redo safe, it is obvious that x behaves as a redo safe variable Note. The recomendation for macro writers is let the user indicate if a redo safe parameter shall be used e.g. follow the pattern the above for macro follows. There is no way to hide ~ from the user currently. It is possible to design the scheme compiler to actually behave nicer and allow macro writers to hide the ~ properties of the code e.g. make the logic behave as no set! has been used. The problem with this approach is that to much need to be changed in scheme to actually dare to propose such an environment. Example 2. Consider backtracking algorithm and that we have three branches f1,f2,f3 of code, we could then just try out for a search like that ( ( f1 f2 f3) f4) if f1 is an infinite tree and ( f1 f4) never succeeds then a solution will never be found if say ( f2 f4) will succeed. Assume that we have a backtracking algorithm utilizing prompts. We could then when ( f1 f4) fails and backtrack, not continue, but store the continuation in a functional queue q and cycle between trying ( f1 f4) ( f2 f4) ... . If we assume that the continuation handling can be made effective we have essentially implemented a logic construct from kanren. Now it is natural to store the queue in variables (fluids can also work). But this is not redo safe. And assuming we would use this in a proof engine interacting with an expert, something is fundamentally wrong. We could use kanren, a beutiful logic engine. But we want to have a dynamic-wind feature as well in the logic engine which is not sound in kanren. Now what you could do when executing is to guard q as a redo safe variable say that we get the continuation k at the toplevel after an abort-to-prompt, then later we could just do (redo-wind-predicate-set! (lambda (x) #t)) and restart the continuation with (k ...) and make sure that we have added (redo-wind-predicate-set! (lambda (x) #f)) after the abort-to-prompt to make sure that we do not restore by misstake. To complexify things you may think of an algorithm where it is included to go through say 100 branches and look for a solution, store the state, go back and collect statistsics and then select the 10 most prommising branches to continue, you may think of doing this including idioms that looks like the special cycling that's described in the beginning of this example. And ontop of that keep an interaction with the user. How do we solve this mess. Keep a parameter I starting with 0, When want to guard a variable you do (with-redo-safe-variables ((a (~ (I)))) code ...) When you introduce a concept at a higher meta level you execute the body with (parameterize ((I (+ (I) 1))) code ...) So the handler get's a level I and the body gets (+ I 1). And then at that level, when you want to not redo the body code (redo-wind-predicate-set! (let ((i (~ (I)))) (lambda (x) (<= x i)))) And when you want to store use (redo-wind-predicate-set! (let ((i (~ (I)))) (lambda (x) (> x i)))) Optimizations and implementations: i) To note, if one instantiates the continuation by copy the stored stack on the scheme stack, and a redo safe variable is never put in a lambda and no continuations is created inside the local function then the optimizer can safly keep the variables unboxed. It will then both behave as redo safe varibles and an optimization. One could even view redo safe variables as a semantic generalization of variables kept unboxed on the local funciton stack although they are mutated. ii) For a full implementation of this concept one can piggy pack ontop of parameter implementations or fluid implementations. In guile the native fluid implementation was mutated to include redo safe variables and redo safe parameters. As a result we could get a speedup of 10x or more compared to using dynamic wind. Therefore we do not expect that effective implementation will be impossible or hard to do. There is one aber though. We demand a dynamic version for tail call's. We have not yet thought that concept through yet. Reference implementation: Assume that we have a complete syntax-case system togeter with a syntax parameters and an ideom (current-syntax-parameter-binding id) We cannot implement the construct fully because we need support from the scheme compiler and cannot lean on a macro system. (define old-call/cc call/cc) (define K (make-parameter #f)) (define *env* '()) (define (storage-to-var k id s setter) (let ((r (assoc (list k id s) *env*))) (if r (begin (setter (cdr r)) #t) #f))) (define (var-to-storage k id s val) (set! *env* (cons (list k id s) val) *env*)) (define (wrap f) (lambda (x) (cond ((eq? x #f) #f) ((eq? x #t) #t) (else (f x))))) (define redo-variables-wind-guard? (make-parameter (wrap (lambda (x) #f)))) (define redo-parameters-wind-guard? (make-parameter (wrap (lambda (x) #f)))) (define undo-store? (make-parameter #f)) (define L (make-parameter '())) (define-syntax with-redo-variables (lambda (x) (syntax-case x () (((s v) ...) code ...) (with-syntax (((w ...) (generate-temporaries #'(s ...)))) #'(let ((first? #t) (id (make-id)) (w v) ...) (dynamic-wind (lambda () (if first? (set! first? #f) (begin (if ((redo-variables-wind-guard?) w) (storage-to-var (K) id 's (lambda (v) (set! s v)))) ...))) (lambda () code ...) (lambda () (if (undo-store?) (L (cond (list id 's s) (L))))))))))) (define-syntax-rules (with-redo-prameters* ((p pp u uu v vv i) ...) code ...) (let ((first? #t) (id (make-id))) (let ((pp p) ... (uu u) ... (vv v) ...) (parameterize ((pp uu) ...) (dynamic-wind (lambda () (if first? (set! first? #f) (begin (if ((redo-parameters-wind-guard?) vv) (storage-to-var k id i (lambda (x) (pp x)))) ...))) (lambda () code ...) (lambda () (if (undo-store?) (L (cond (list id i (pp)) (L))))))))))) (define-syntax with-redo-parameters (lambda (x) (syntax-case x () ((_ ((p u v) ...) code ...) (with-syntax ((pp (generate-temporaries #'(p ...))) (uu (generate-temporaries #'(p ...))) (vv (generate-temporaries #'(p ...))) (i (iota (length #'(p ...))))) #'(with-redo-parameters* ((pp p uu u vv v o) ...) code ...)))))) ;;Applying guile's abort-to-prompt might not work but you get the point (define-syntax-rule (abort-to-prompt~ tag . l) (apply abort-to-prompt tag (let ((ret (list . l))) (L '()) (undo-store? #t) ret))) (define-syntax-rule (call-with-prompt~ tag thunk handler) (call-with-prompt tag thunk (lambda (k . v) (if (undo-store?) (begin (for-each (lambda (x) (call-with-values (lambda () (apply values x)) (lambda (id s v) (var-to-storage k id s (v))))) (L)) (L '()) (undo-store? #f) (apply handler (lambda x (parameterize ((K k)) (apply k x))) v))))))