Lazy loading of opaque proofs: fast as -dont-load-proofs without its drawbacks

 The recent experiment with -dont-load-proofs in the stdlib showed that
 this options isn't fully safe: some axioms were generated (Include ?
 functor application ? This is still to be fully understood).

 Instead, I've implemented an idea of Yann: only load opaque proofs when
 we need them. This is almost as fast as -dont-load-proofs (on the stdlib,
 we're now 15% faster than before instead of 20% faster with -dont-load-proofs),
 but fully compatible with Coq standard behavior.

 Technically, the const_body field of Declarations.constant_body now regroup
 const_body + const_opaque + const_inline in a ternary type. It is now either:
  - Undef : an axiom or parameter, with an inline info
  - Def : a transparent definition, with a constr_substituted
  - OpaqueDef : an opaque definition, with a lazy constr_substitued
 Accessing the lazy constr of an OpaqueDef might trigger the read on disk of
 the final section of a .vo, where opaque proofs are located.

 Some functions (body_of_constant, is_opaque, constant_has_body) emulate
 the behavior of the old fields. The rest of Coq (including the checker)
 has been adapted accordingly, either via direct access to the new const_body
 or via these new functions. Many places look nicer now (ok, subjective notion).

 There are now three options: -lazy-load-proofs (default), -force-load-proofs
 (earlier semantics), -dont-load-proofs. Note that -outputstate now implies
 -force-load-proofs (otherwise the marshaling fails on some delayed lazy).

 On the way, I fixed what looked like a bug : a module type
 (T with Definition x := c) was accepted even when x in T was opaque.
 I also tried to clarify Subtyping.check_constant.

git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@13952 85f007b7-540e-0410-9357-904b9bb8a0f7

1 files changed, 29 insertions, 5 deletions

diff --git a/checker/declarations.mli b/checker/declarations.mli
index b5038996c..90beb326b 100644
--- a/checker/declarations.mli
+++ b/checker/declarations.mli
@@ -29,16 +29,40 @@ type constr_substituted
 val force_constr : constr_substituted -> constr
 val from_val : constr -> constr_substituted
 
-type inline = int option (* inlining level, None for no inlining *)
+(** Beware! In .vo files, lazy_constr are stored as integers
+   used as indexes for a separate table. The actual lazy_constr is restored
+   later, by [Safe_typing.LightenLibrary.load]. This allows us
+   to use here a different definition of lazy_constr than coqtop:
+   since the checker will inspect all proofs parts, even opaque
+   ones, no need to use Lazy.t here *)
+
+type lazy_constr
+val force_lazy_constr : lazy_constr -> constr
+val lazy_constr_from_val : constr_substituted -> lazy_constr
+
+(** Inlining level of parameters at functor applications.
+    This is ignored by the checker. *)
+
+type inline = int option
+
+(** A constant can have no body (axiom/parameter), or a
+    transparent body, or an opaque one *)
+
+type constant_def =
+  | Undef of inline
+  | Def of constr_substituted
+  | OpaqueDef of lazy_constr
 
 type constant_body = {
     const_hyps : section_context; (* New: younger hyp at top *)
-    const_body : constr_substituted option;
+    const_body : constant_def;
     const_type : constant_type;
     const_body_code : to_patch_substituted;
-    const_constraints : Univ.constraints;
-    const_opaque : bool;
-    const_inline : inline }
+    const_constraints : Univ.constraints }
+
+val body_of_constant : constant_body -> constr_substituted option
+val constant_has_body : constant_body -> bool
+val is_opaque : constant_body -> bool
 
 (* Mutual inductives *)