HUGE COMMIT

1. when applying a functor F(X) := B to a module M, the obtained module
   is no longer B{X.t := M.t for all t}, but B{X.t := b where b is the
   body of t in M}. In principle it is now easy to fine tune the behaviour
   to choose whether b or M.t must be used. This change implies modifications
   both inside and outside the kernel.
2. for each object in the library it is now necessary to define the behaviour
   w.r.t. the substitution {X.t := b}. Notice that in many many cases the
   pre-existing behaviour w.r.t. the substitution {X.t := M.t} was broken
   (in the sense that it used to break several invariants). This commit
   fixes the behaviours for most of the objects, excluded
     a) coercions: a future commit should allow any term to be declared
        as a coercion; moreover the invariant that just a coercion path
        exists between two classes will be broken by the instantiation.
     b) global references when used as arguments of a few tactics/commands
   In all the other cases the behaviour implemented is the one that looks
   to me as the one expected by the user (if possible):
   [ terminology: not expanded (X.t := M.t) vs expanded (X.t := b) ]
     a) argument scopes: not expanded
     b) SYNTAXCONSTANT: expanded
     c) implicit arguments: not expanded
     d) coercions: expansion to be done (for now not expanded)
     e) concrete syntax tree for patterns: expanded
     f) concrete syntax tree for raw terms: expanded
     g) evaluable references (used by unfold, delta expansion, etc.): not
        expanded
     h) auto's hints: expanded when possible (i.e. when the expansion of the
        head of the pattern is rigid)
     i) realizers (for program extraction): nothing is done since the actual
        code does not look for the realizer of definitions with a body;
        however this solution is fragile.
     l) syntax and notation: expanded
     m) structures and canonical structures: an invariant says that no
        parameter can happear in them ==> the substitution always yelds the
        original term
     n) stuff related to V7 syntax: since this part of the code is doomed
        to disappear, I have made no effort to fix a reasonable semantics;
        not expanded is the default one applied
     o) RefArgTypes: to be understood. For now a warning is issued whether
        expanded != not expanded, and the not expanded solution is chosen.


git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@6555 85f007b7-540e-0410-9357-904b9bb8a0f7

1 files changed, 10 insertions, 2 deletions

diff --git a/contrib/extraction/table.ml b/contrib/extraction/table.ml
index f95f81ee9..ce2a99c6d 100644
--- a/contrib/extraction/table.ml
+++ b/contrib/extraction/table.ml
@@ -337,14 +337,22 @@ let add_inline_entries b l =
 
 (* Registration of operations for rollback. *)
 
-let (inline_extraction,_) = 
+let (inline_extraction,_) =
   declare_object 
     {(default_object "Extraction Inline") with 
        cache_function = (fun (_,(b,l)) -> add_inline_entries b l);
        load_function = (fun _ (_,(b,l)) -> add_inline_entries b l);
        export_function = (fun x -> Some x); 
        classify_function = (fun (_,o) -> Substitute o); 
-       subst_function = (fun (_,s,(b,l)) -> (b,(List.map (subst_global s) l))) }
+       (*CSC: The following substitution may istantiate a realized parameter.
+         The right solution would be to make the substitution erase the
+         realizer from the table. However, this is not allowed by Coq.
+         In this particular case, though, keeping the realizer is place seems
+         to be harmless since the current code looks for a realizer only
+         when the constant is a parameter. However, if this behaviour changes
+         subtle bugs can happear in the future. *)
+       subst_function =
+        (fun (_,s,(b,l)) -> (b,(List.map (fun x -> fst (subst_global s x)) l)))}
 
 let _ = declare_summary "Extraction Inline"
 	  { freeze_function = (fun () -> !inline_table);