Commit 962e6823 authored by POTTIER Francois's avatar POTTIER Francois Committed by POTTIER Francois
Browse files

Fix the safe-expansion check as suggested by Yann in merge request !8.

Add a test that must be rejected because expansion diverges,
and was not properly rejected until now.
parent 0adb78e6
# Changes
## 2020/03/04
* Fix the static check that is performed prior to expanding the parameterized
nonterminal symbols. The previous check was incorrect: it would sometimes
accept a grammar whose expansion would diverge.
## 2020/02/11
* Re-implement Menhir's default algorithm for constructing LR(1) automata,
......@@ -73,23 +73,54 @@ let formals : formal array =
there is a "dangerous" edge from the formal parameter F/i to the formal G/j.
This reflects the fact that there is a flow from F/i to G/j. This flow is
"dangerous" in the sense that it is size-increasing: X is transformed to
H(..., X, ...). *)
H(..., X, ...). In this example, there should also be a safe edge from the
formal F/i to the formal H/k.
More generally, an occurrence of X in the right-hand side can occur deeply
nested in a context K:
F(..., X, ...):
... K[X] ...
In that case, we must create as many edges as the context K is deep, and
all of these edges should be dangerous, except the one that corresponds
to the innermost layer of K, which is safe.
Another way of putting this is, when the left-hand side is:
F(..., X, ...): # where X is the i-th formal parameter of F
and the right-hand side contains a possibly-nested occurrence of:
G(..., K[X], ...) # where K[X] is the j-th actual parameter of G
then we must create an edge of F/i to G/j, and this edge is safe if and
only if the context K is empty, i.e., X occurs at depth 0 in K[x]. *)
(* The code below has quadratic complexity because [Parameters.occurs_deep]
is expensive. In principle, we could achieve linear complexity by first
annotating subterm (bottom-up) with a Boolean flag that indicates whether
[x] occurs (shallowly/deeply) in it; that would allow us to implement
[occurs_deep] in constant time. However, in practice, quadratic complexity
is probably good enough. *)
type edge =
| Safe
| Dangerous
let successors_parameter (f : edge -> formal -> unit) x (param : parameter) =
let rec successors_parameter (f : edge -> formal -> unit) x (param : parameter) =
match param with
| ParameterVar _ ->
(* This is not an application. No successors. *)
| ParameterApp (sym, params) ->
let nt = value sym in
List.iteri (fun i param ->
(* Check, recursively, the applications that appear inside [param]. *)
successors_parameter f x param;
(* If [x] occurs in the [i]-th actual parameter of this application,
then there is an edge to the formal [nt, i]. Whether it is a safe
or dangerous edge depends on whether [x] occurs shallow or deep. *)
List.iteri (fun i param ->
if Parameters.occurs_shallow x param then
f Safe (nt, i)
else if Parameters.occurs_deep x param then
Error: the parameterized nonterminal symbols in this grammar
cannot be expanded away: expansion would not terminate.
The first formal parameter of "seq(_)" grows without bound.
(* This grammar must be rejected, as the expansion of parameterized
nonterminal symbols does not terminate: seq(t) calls seq(wrap(t)).
This problem has been reported by Stéphane Lescuyer
and studied by Yann Régis-Gianas.
%token A
seq(A) {}
t {}
| wrap(seq(wrap(t))) {}
......@@ -8133,6 +8133,14 @@
(rule (alias expansion-capture)
(action (diff ../bad/expansion-capture.exp expansion-capture.out)))
(rule (target expansion_diverges.out) (deps ../bad/expansion_diverges.mly)
(with-outputs-to expansion_diverges.out
(chdir ../bad (with-accepted-exit-codes (not 0) (run menhir %{deps}))))))
(rule (alias expansion_diverges)
(action (diff ../bad/expansion_diverges.exp expansion_diverges.out)))
(rule (target expression.out) (deps ../bad/expression.mly)
(with-outputs-to expression.out
......@@ -9421,6 +9429,7 @@
(alias error-reserved)
(alias execparser)
(alias expansion-capture)
(alias expansion_diverges)
(alias expression)
(alias expression-expected)
(alias expressions)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment