Re: [racket-dev] A proposal for parametric opaque types in Typed Racket
On Jan 30, 2015, at 3:59 PM, Alexis King lexi.lam...@gmail.com wrote: No, the typechecker can’t make any assumptions about the results of opaque types. If you explicitly instantiate a Posn with the type Real, the typechecker should only guarantee the result will be Real. Annotate the type as (U 1 2), though, and obviously it would need to ensure that remains invariant. How about this program: untyped.rkt: #lang racket (provide (all-defined-out)) (define (make-posn x y) (list 3 y)) ; bad typed.rkt: #lang typed/racket ; make Posn parametric (define-type (Posn X Y) (List X Y)) (require/typed untyped.rkt [make-posn (All (X Y) X Y - (Posn X Y))]) (: p : (Posn Real Real)) (define p (make-posn 1 2)) This gives this error: . . make-posn: broke its contract promised: X3 produced: 3 in: the car of the range of (parametric-/c (X3 Y4) (-* (X3 Y4) () (values (cons/c X3 (cons/c Y4 g6) contract from: (interface for make-posn) blaming: (interface for make-posn) (assuming the contract is correct) at: …./typed.rkt:5.16 I think it’s a good thing that it checks that it actually gives you the value that you gave it, and not just something like 3 even if it happens to match the type you want. And I think parametric opaque types should behave in a similar way, and to do that you would need the opaque value to be wrapped in another opaque structure, which would store either the contracts or the set of values that would pass the contracts or something like that. On Jan 30, 2015, at 12:30, Alexander D. Knauth alexan...@knauth.org wrote: On Jan 30, 2015, at 1:53 PM, Alexis King lexi.lam...@gmail.com wrote: No, it doesn’t need to be wrapped in an opaque structure. Wrapping it in an opaque structure would add a layer of indirection for absolutely no gain. Remember, the value itself is already, by definition, opaque. The only way typed code can manipulate the value is by passing it to other functions imported via require/typed. This means that contracts only need to be generated wherever those functions are called. This can be done without wrapping or unwrapping anything because all the information required to generate those contracts is known at expansion-time. The typechecker simply needs to insert the relevant contracts at the relevant locations. Imagine a program like this: #lang typed/racket (require typed/lang/posn) (: p : (Posn Real Real)) ; I’m assuming Posn is parametric over 2 tvars, not 1 (define p (posn 1 2)) (: x : Real) (define x (posn-x p)) As far as the type checker would check, it would check that the result of posn-x is a Real, but I think that the runtime contract it should also check that it returns 1, because posn could have been instantiated as (Posn 1 2). #lang typed/racket (require typed/lang/posn/mutable) ; like typed/lang/posn, but providing mutation too (: p : (Posn Real Real)) (define p (posn 1 2)) (: x : Real) (define x (posn-x p)) (set-posn-x! p 3) (: x2 : Real) (define x2 (posn-x p)) Here, even though the type checker only cares that it’s a number, it should check that x2 definition returns either 1 or 3, since both were provided as x values for the posn p. For it to keep track of these at runtime, (and it would have to be runtime) the contracts would have to be with the actual posn value in an opaque structure, which would have contracts sort of like (new-∀/c) that would check these things, although I don’t think it would have to wrap the inner values, but just record them so that when posn-x is called on one of these things, it checks that it was one of the values that was passed in to either a constructor or setter function. On Jan 30, 2015, at 07:27, Alexander D. Knauth alexan...@knauth.org wrote: On Thu, Jan 29, 2015, at 09:03 PM, Alexis King wrote: It isn’t wrapped in an opaque structure. That wasn’t a part of my proposal, and while I didn’t think of it until you brought it up, I still think it’s unnecessary and doesn’t add any convenience. I think the opaque structures would be necessary for the kind of sharing wrappers between functions that you describe just before section 2.1, except that instead of the sub-values being wrapped on the untyped side, the whole thing is wrapped on the typed side, and there is a contract that wraps it and unwraps it when it goes from untyped to typed and back. For parametric types, they have to also work if the type was constrained to the exact set of values that were provided, which means that if you provide two numbers, say 1 and 2, it has to return a posn with not just any two numbers, but values of the type (U 1 2), since A could have been constrained to (U 1 2). So it has to be wrapped somehow, and I think wrapping it on the typed side makes more sense. Perhaps I’m not understanding you properly, but your “one-length string”
Re: [racket-dev] A proposal for parametric opaque types in Typed Racket
No, it doesn’t need to be wrapped in an opaque structure. Wrapping it in an opaque structure would add a layer of indirection for absolutely no gain. Remember, the value itself is already, by definition, opaque. The only way typed code can manipulate the value is by passing it to other functions imported via require/typed. This means that contracts only need to be generated wherever those functions are called. This can be done without wrapping or unwrapping anything because all the information required to generate those contracts is known at expansion-time. The typechecker simply needs to insert the relevant contracts at the relevant locations. On Jan 30, 2015, at 07:27, Alexander D. Knauth alexan...@knauth.org wrote: On Thu, Jan 29, 2015, at 09:03 PM, Alexis King wrote: It isn’t wrapped in an opaque structure. That wasn’t a part of my proposal, and while I didn’t think of it until you brought it up, I still think it’s unnecessary and doesn’t add any convenience. I think the opaque structures would be necessary for the kind of sharing wrappers between functions that you describe just before section 2.1, except that instead of the sub-values being wrapped on the untyped side, the whole thing is wrapped on the typed side, and there is a contract that wraps it and unwraps it when it goes from untyped to typed and back. For parametric types, they have to also work if the type was constrained to the exact set of values that were provided, which means that if you provide two numbers, say 1 and 2, it has to return a posn with not just any two numbers, but values of the type (U 1 2), since A could have been constrained to (U 1 2). So it has to be wrapped somehow, and I think wrapping it on the typed side makes more sense. Perhaps I’m not understanding you properly, but your “one-length string” idea sounds like it has little to do with this opaque type problem and more to do with the fact that you want refinement types in Typed Racket. I do, too! But I don’t think hacking the opaque type system is going to help you with that. Well, yeah, refinement types would be the real solution for this particular example, but if I do want to constrain it to strings of length 1, opaque types are the only option for now, and they actually work fine. My point was you couldn't do this type of thing with the opaque structures and you would probably get weird errors if you tried. (See below because there might be a solution?) (Also, as for the box example, I’m actually a little surprised that doesn’t contract error. Seems like a bug to me, but perhaps I’m missing some idiosyncrasies of the type system. Either way, it’s precisely that kind of craziness I was referring to when I compared casting parametric opaque types to casting mutable types.) There is a bug report for it here, and the solution proposed by Sam Tobin-Hochstadt would be for cast to generate 2 contracts, one for the original type, one for the new type, but that never got implemented. http://bugs.racket-lang.org/query/?cmd=viewpr=13626 http://bugs.racket-lang.org/query/?cmd=viewpr=13626 Actually now that I think about it the two-contract solution might be able to solve the previous problem, since the original contract could unwrap the value before it is passed to the new contract? I'm not sure though. The value inside the cast would be from the typed side, then it is passed through the orig contract as if it were going to the typed side, which would unwrap it, and then that unwrapped value would be passed to the new contract as if it were flowing from the untyped side to the typed side. On Jan 29, 2015, at 20:50, Alexander D. Knauth alexan...@knauth.org mailto:alexan...@knauth.org wrote: On Jan 29, 2015, at 11:34 PM, Alexis King lexi.lam...@gmail.com mailto:lexi.lam...@gmail.com wrote: But the problem is that if it’s an opaque type then it can’t unwrap it once the value is returned from make-posn. Yes, that’s precisely the problem. Your point about implementing everything as single-valued structs on the typed side is an interesting one, though I don’t think it ultimately solves any problems. The fact that the typed side knowsnothingabout the contents of the value is what makes this such a tricky problem. As for this: But then you couldn’t do any operations on it except those that you use import with require/typed, right? That’s completely correct. That’s why it’s “opaque.” And what happens if you use cast on one of these things? That’s a little more interesting. Usingcaston an object of this type would never fail (unless, of course, it didn’t actually satisfy the basicposn?predicate), but it would possibly introduce failures in the future since it would affect the contracts generated forposn-xandposn-y, for example. To make that more clear, casting a(Posn Real)to a(Posn
Re: [racket-dev] A proposal for parametric opaque types in Typed Racket
On Jan 30, 2015, at 1:53 PM, Alexis King lexi.lam...@gmail.com wrote: No, it doesn’t need to be wrapped in an opaque structure. Wrapping it in an opaque structure would add a layer of indirection for absolutely no gain. Remember, the value itself is already, by definition, opaque. The only way typed code can manipulate the value is by passing it to other functions imported via require/typed. This means that contracts only need to be generated wherever those functions are called. This can be done without wrapping or unwrapping anything because all the information required to generate those contracts is known at expansion-time. The typechecker simply needs to insert the relevant contracts at the relevant locations. Imagine a program like this: #lang typed/racket (require typed/lang/posn) (: p : (Posn Real Real)) ; I’m assuming Posn is parametric over 2 tvars, not 1 (define p (posn 1 2)) (: x : Real) (define x (posn-x p)) As far as the type checker would check, it would check that the result of posn-x is a Real, but I think that the runtime contract it should also check that it returns 1, because posn could have been instantiated as (Posn 1 2). #lang typed/racket (require typed/lang/posn/mutable) ; like typed/lang/posn, but providing mutation too (: p : (Posn Real Real)) (define p (posn 1 2)) (: x : Real) (define x (posn-x p)) (set-posn-x! p 3) (: x2 : Real) (define x2 (posn-x p)) Here, even though the type checker only cares that it’s a number, it should check that x2 definition returns either 1 or 3, since both were provided as x values for the posn p. For it to keep track of these at runtime, (and it would have to be runtime) the contracts would have to be with the actual posn value in an opaque structure, which would have contracts sort of like (new-∀/c) that would check these things, although I don’t think it would have to wrap the inner values, but just record them so that when posn-x is called on one of these things, it checks that it was one of the values that was passed in to either a constructor or setter function. On Jan 30, 2015, at 07:27, Alexander D. Knauth alexan...@knauth.org wrote: On Thu, Jan 29, 2015, at 09:03 PM, Alexis King wrote: It isn’t wrapped in an opaque structure. That wasn’t a part of my proposal, and while I didn’t think of it until you brought it up, I still think it’s unnecessary and doesn’t add any convenience. I think the opaque structures would be necessary for the kind of sharing wrappers between functions that you describe just before section 2.1, except that instead of the sub-values being wrapped on the untyped side, the whole thing is wrapped on the typed side, and there is a contract that wraps it and unwraps it when it goes from untyped to typed and back. For parametric types, they have to also work if the type was constrained to the exact set of values that were provided, which means that if you provide two numbers, say 1 and 2, it has to return a posn with not just any two numbers, but values of the type (U 1 2), since A could have been constrained to (U 1 2). So it has to be wrapped somehow, and I think wrapping it on the typed side makes more sense. Perhaps I’m not understanding you properly, but your “one-length string” idea sounds like it has little to do with this opaque type problem and more to do with the fact that you want refinement types in Typed Racket. I do, too! But I don’t think hacking the opaque type system is going to help you with that. Well, yeah, refinement types would be the real solution for this particular example, but if I do want to constrain it to strings of length 1, opaque types are the only option for now, and they actually work fine. My point was you couldn't do this type of thing with the opaque structures and you would probably get weird errors if you tried. (See below because there might be a solution?) (Also, as for the box example, I’m actually a little surprised that doesn’t contract error. Seems like a bug to me, but perhaps I’m missing some idiosyncrasies of the type system. Either way, it’s precisely that kind of craziness I was referring to when I compared casting parametric opaque types to casting mutable types.) There is a bug report for it here, and the solution proposed by Sam Tobin-Hochstadt would be for cast to generate 2 contracts, one for the original type, one for the new type, but that never got implemented. http://bugs.racket-lang.org/query/?cmd=viewpr=13626 Actually now that I think about it the two-contract solution might be able to solve the previous problem, since the original contract could unwrap the value before it is passed to the new contract? I'm not sure though. The value inside the cast would be from the typed side, then it is passed through the orig contract as if it were going to the typed side, This was a typo, I meant to say “as if it were going to the
Re: [racket-dev] [plt] Push #29677: master branch updated
This change seemed to change the format of .dep files, likely as intended
to add the indirect dependencies. Is there any documentation of what the
format is supposed to be? Currently I've just been trying to read cm.rkt
and understand how it treats them.
On Thu, Jan 8, 2015 at 9:31 AM, mfl...@racket-lang.org wrote:
mflatt has updated `master' from c56c9250f1 to 95e85ec5bd.
http://git.racket-lang.org/plt/c56c9250f1..95e85ec5bd
=[ 2 Commits ]==
Directory summary:
45.1% pkgs/racket-doc/scribblings/raco/
4.7% pkgs/racket-doc/scribblings/reference/
47.5% racket/collects/compiler/
~~
fe9a04d Matthew Flatt mfl...@racket-lang.org 2015-01-08 09:11
:
| doc tweaks for `raco {setup,make}`
:
M pkgs/racket-doc/scribblings/raco/make.scrbl | 4 ++--
M pkgs/racket-doc/scribblings/raco/setup.scrbl | 22
--
~~
95e85ec Matthew Flatt mfl...@racket-lang.org 2015-01-08 09:57
:
| add support for indirect CM dependencies; use in `lazy-require`
|
| If module M in package P imports module N from package Q,
| and if N has a `lazy-require` for a module in R that is
| triggered during the compilation of M, then P doesn't really
| depend on R; P depends on Q, and Q depends on R, and P
| shoudn't necessarily know anything about Q. At the same time,
| a change to the file in R means that M must be recompiled.
| So, continue to track the compilation dependency, but mark
| it as indirect so that the package-dependency checker can
| ignore the dependency.
:
M pkgs/racket-doc/scribblings/raco/make.scrbl | 33 -
M racket/collects/compiler/cm-accomplice.rkt| 14 +++---
M racket/collects/compiler/cm.rkt | 49
++--
M racket/collects/racket/lazy-require.rkt | 2 +-
M racket/collects/setup/private/pkg-deps.rkt| 1 +
M .../racket-doc/scribblings/reference/syntax.scrbl | 8 ++--
=[ Overall Diff ]===
pkgs/racket-doc/scribblings/raco/make.scrbl
~~~
--- OLD/pkgs/racket-doc/scribblings/raco/make.scrbl
+++ NEW/pkgs/racket-doc/scribblings/raco/make.scrbl
@@ -123,7 +123,7 @@ would create only @filepath{compiled/b_rkt.zo} and
@; --
-@section{Dependency Files}
+@section[#:tag Dependency Files]{Dependency Files}
In addition to a bytecode file, @exec{raco make} creates a file
@filepath{compiled/@nonterm{name}_@nonterm{ext}.dep} that records
@@ -538,7 +538,7 @@ messages are instances of a
@racket[parallel-compile-event] prefab structure:
@racketblock[
(struct parallel-compile-event (worker event) #:prefab)
-].
+]
The worker field is the index of the worker that the created the event.
The event
field is a @racket[compile-event] as document in
@@ -550,25 +550,36 @@ field is a @racket[compile-event] as document in
@defmodule[compiler/cm-accomplice]
-@defproc[(register-external-file [file (and path? complete-path?)])
void?]{
+@defproc[(register-external-file [file (and path? complete-path?)]
+ [#:indirect? indirect? any/c #f])
+ void?]{
-Logs a message (see @racket[log-message]) at level @racket['info] to
-a logger named @racket['cm-accomplice]. The
-message data is a @racketidfont{file-dependency} prefab structure type
-with two fields; the first field's value is @racket[file] and the second
-field's value is @racket[#f] (to indicate a non-module dependency).
+Logs a message (see @racket[log-message]) at level @racket['info] to a
+logger named @racket['cm-accomplice]. The message data is a
+@racketidfont{file-dependency} prefab structure type with two fields;
+the first field's value is @racket[file] and the second field's value
+is @racket[#f] (to indicate a non-module dependency). If the
+@racket[indirect?] argument is true, the data is more specifically an
+instance of a @racketidfont{file-dependency/indirect} prefab structure
+type that is a subtype of @racketidfont{file-dependency} with no new
+fields.
A compilation manager implemented by @racketmodname[compiler/cm] looks
-for such messages to register an external dependency. The compilation
-manager records (in a @filepath{.dep} file) the path as contributing
-to the implementation of the module currently being
+for such messages to register an external dependency. In response, the
+compilation manager records (in a @filepath{.dep} file) the path as
+contributing to the implementation of the module currently being
compiled. Afterward, if the registered file is modified, the
-compilation manager will know to recompile the module.
+compilation manager will know to recompile the module. An ``indirect''
+dependency has no effect on recompilation, but it can signal to other
+tools, such as a package-dependency
Re: [racket-dev] feature request: thread-safe memoize-evt
Thanks for your time. On Thu, 2015-01-29 at 12:55 -0700, Matthew Flatt wrote: Would the simpler `once-evt` work in your situation, or do you need the guarantee that only one wait of E happens at a time? OK, my original goal is to implement a remote method call multiplexer. The kind where you send a message with an identifier and the remote party eventually replies with the same identifier or possibly with a well-known one to indicate an out-of-bound notification. So, I'd like to end up with (something-receive-evt something key) and (something-call-evt something key request). I can see two ways to do that; 1. Have a dispatch table with pending messages. 2. Broadcast all incoming messages to all waiters. Dispatch table requires non-racy invalidation of pending requests to prevent resource leaks. Broadcasting messages requires construction of a thread-safe memoizing event combinator (produces next-evt and a single result). Also, I have a feeling that I could get close to (1) if I keep some cleanup thread and utilize nack-guard-evt, but something tells me it might be racy. Best regards, Jan Dvorak _ Racket Developers list: http://lists.racket-lang.org/dev
