Merge pull request #8 from jsmaniac/fix-use-site-scope

Remove the extra use-site scope on the body of define-?-expander.

private/define-expanders.rkt

@@ -7,6 +7,13 @@
 
 (provide define-expander-type)
 
+(define-for-syntax (remove-use-site-scope stx)
+  (define bd
+    (syntax-local-identifier-as-binding (syntax-local-introduce #'here)))
+  (define delta
+    (make-syntax-delta-introducer (syntax-local-introduce #'here) bd))
+  (delta stx 'remove))
+
 (define-syntax define-expander-type
   (syntax-parser
     [(_ name:id)
@@ -15,13 +22,16 @@
                                [?-expander? "~a-expander?"]
                                [define-?-expander "define-~a-expander"]
                                [expand-all-?-expanders "expand-all-~a-expanders"])
-                       #'(begin
+                       #`(begin
                            (define-for-syntax ?-expander-type (make-expander-type))
                            (define-for-syntax (make-?-expander transformer)
                              (expander ?-expander-type transformer))
                            (define-for-syntax (?-expander? v)
                              (expander-of-type? ?-expander-type v))
-                           (define-syntax-rule (define-?-expander expander-name transformer)
-                             (define-syntax expander-name (make-?-expander transformer)))
+                           (define-syntax define-?-expander
+                             (syntax-parser
+                               [(_ expander-name:id transformer:expr)
+                                (remove-use-site-scope
+                                 #'(define-syntax expander-name (make-?-expander transformer)))]))
                            (define-for-syntax (expand-all-?-expanders stx)
                              (expand-syntax-tree-with-expanders-of-type ?-expander-type stx))))]))

test/test-define-x-expander-use-site-scope-simple.rkt

@@ -0,0 +1,175 @@
+#lang racket
+
+;; For a more realistic use-case justifying why this behaviour matters, see the
+;; discussion at https://github.com/jackfirth/generic-syntax-expanders/pull/8
+;; and in particular the test file at
+;;   https://github.com/jsmaniac/generic-syntax-expanders
+;;     /blob/ec43791028715221c678f8536389a39ee760ed98
+;;     /test/test-define-x-expander-use-site-scope.rkt
+
+(require generic-syntax-expanders
+         rackunit)
+
+(define-expander-type foo)
+
+(define-syntax (expand-foos stx)
+  (syntax-case stx ()
+    [(_ body)
+     (expand-all-foo-expanders #'body)]))
+
+;; Without PR #8, the `x` in `bb` is not `bound-identifier=?` to the `x`
+;; defined by `aa`, despite the use of `syntax-local-introduce` to make the
+;; macros unhygienic.
+;;
+;; This happens because `define-foo-expander` added an extra "use-site" scope
+;; to the body of `aa` and a different "use-site" scope to the body of `bb`.
+(define-foo-expander aa
+  (λ (_)
+    (syntax-local-introduce #'[x #t])))
+
+(define-foo-expander bb
+  (λ (_)
+    (syntax-local-introduce #'x)))
+
+;; Due to the way `let` itself adds scopes to its definition and body, this
+;; makes the identifiers `x` from `aa` and `x` from `bb` distinct, and the
+;; latter cannot be used to refer to the former.
+;;
+;; Approximately, the code below expands to
+;;
+;;     (let [(x⁰¹ #t)]
+;;       x⁰²)
+;;
+;; The `let` form then adds a "local" scope to both occurrences of `x`, and an
+;; internal-definition context "intdef" scope to the `x` present in the body of
+;; the `let` (but not to the one present in the bindings). The expanded form
+;; therefore becomes:
+;;
+;;     (let [(x⁰¹³ #t)]
+;;       x⁰²³⁴)
+;;
+;; where:
+;;   ⁰ are the module's scopes
+;;   ¹ is the undesired "use-site" scope added by `define-foo-expander` on `aa`
+;;   ² is the undesired "use-site" scope added by `define-foo-expander` on `bb`
+;;   ³ is the "local" scope added by `let`
+;;   ⁴ is the "intdef" scope added by `let`
+;;
+;; Since {0,2,3,4} is not a subset of {0,1,3}, the `x` inside the `let` is
+;; unbound.
+(test-true
+ "Test that `x` as produced by `(bb)` is correctly bound by the `x`
+ introduced by `(aa)`.
+
+This test fails without the PR #8 patch, because the body of `bb` and the body
+ of `aa` each have a different use-site scope, introduced by accident by
+ `define-foo-expander`. The occurrence of `x` introduced by `aa` and the
+ occurrence of `x` introduced by `bb` therefore have different scopes, and the
+ latter is not bound by the former.
+
+Without the PR #8 patch, this test case will not compile, and will fail with
+ the error `x: unbound identifier in module`."
+ (expand-foos
+  (let ((aa))
+    (bb))))
+
+;; ----------
+
+;; It is worth noting that `define` seems to strip the "use-site" scopes present
+;; on the defined identifier. If the code above is changed so that a `define`
+;; form is used, the problem does not occur:
+(define-foo-expander aa-def
+  (λ (_)
+    (syntax-local-introduce #'[define y #t])))
+
+(define-foo-expander bb-def
+  (λ (_)
+    (syntax-local-introduce #'y)))
+
+;; This is because the code below expands to:
+;;
+;;     (begin (define y⁰ #t)
+;;            (define y-copy y⁰²⁵)
+;;
+;; where:
+;;   ⁰ are the module's scopes
+;;   ¹ is the undesired "use-site" scope added by `define-foo-expander` on `aa`
+;;     and it is stripped by `define` from the first `y`
+;;   ² is the undesired "use-site" scope added by `define-foo-expander` on `bb`
+;;   ⁵ is the "use-site" scope added because it is in an expression position
+;;
+;; Since {0,2,5} is a subset of {0}, the second `y` refers to the first `y`.
+(expand-foos
+ (begin (aa-def)
+        (define y-copy (bb-def))))
+(test-true
+ "Test that `y` as produced by `(bb-def)` is correctly bound by the `y`
+ defined by `(aa-def)`.
+
+This test succeeds without the PR #8 patch, which shows that `define` removes
+ all use-site scopes on the defined identifier (or at least it removes all the
+ use-site scopes present in this example). This can be checked in the macro
+ debugger, and explains why the test case did not fail with a simple `define`,
+ but does fail with a binding introduced by a `let`."
+ y-copy)
+
+;; ----------
+
+;; The code below attempts to remove the extra "use-site" scope with
+;; `syntax-local-identifier-as-binding`. However, that function does
+;; not remove all use-site scopes, unlike the `define` above.
+
+(define-foo-expander aa-as-binding
+  (λ (_)
+    #`[#,(syntax-local-identifier-as-binding (syntax-local-introduce #'z)) #t]))
+
+(define-foo-expander bb-as-binding
+  (λ (_)
+    (syntax-local-introduce #'z)))
+
+(test-true
+ "Test that `z` as produced by `(bb-as-binding)` is correctly bound by
+ the `z` defined by `(aa-as-binding)`.
+
+This test fails without the PR #8 patch, which shows that that unlike `define`,
+ the `syntax-local-identifier-as-binding` function does not remove all use-site
+ scopes.
+
+Without the PR #8 patch, this test case will not compile, and will fail with
+ the error `z: unbound identifier in module`."
+ (expand-foos
+  (let ((aa-as-binding))
+    (bb-as-binding))))
+
+;; ----------
+
+;; The `cc` expander acts either as aa or as bb depending on the keyword passed
+;; to it. Without PR #8, the code below still compiles fine.
+;;
+;;
+;; The fact that it worked without the patch testifies that the extra scope
+;;   was added on the definition of `aa` and `bb`, instead of being a new fresh
+;;   scope added each time the expander is called. Here, we have two calls to
+;;   the `cc` expander successfully communicating via the `w` variable, thanks
+;;   to `syntax-local-introduce` (which makes the macros unhygienic).
+(define-foo-expander cc
+  (λ (stx)
+    (syntax-case stx ()
+      [(_ #:aa)
+       (begin
+         (syntax-local-introduce #'[w #t]))]
+      [(_ #:bb)
+       (begin
+         (syntax-local-introduce #'w))])))
+
+(test-true
+ "Test that `w` as produced by `(cc #:bb)` is correctly bound by the `w`
+ introduced by `(cc #:aa)`.
+
+This test succeeds without the PR #8 patch, which shows that the extra
+ scopes are per-expander and not per-invocation. Expanders can still be
+ unhygienic using `syntax-local-introduce`, but can communicate only with
+ themselves."
+ (expand-foos
+  (let ((cc #:aa))
+    (cc #:bb))))