diff options
Diffstat (limited to 'contrib/extraction/mlutil.ml')
-rw-r--r-- | contrib/extraction/mlutil.ml | 1136 |
1 files changed, 1136 insertions, 0 deletions
diff --git a/contrib/extraction/mlutil.ml b/contrib/extraction/mlutil.ml new file mode 100644 index 00000000..fbe423a7 --- /dev/null +++ b/contrib/extraction/mlutil.ml @@ -0,0 +1,1136 @@ +(************************************************************************) +(* v * The Coq Proof Assistant / The Coq Development Team *) +(* <O___,, * CNRS-Ecole Polytechnique-INRIA Futurs-Universite Paris Sud *) +(* \VV/ **************************************************************) +(* // * This file is distributed under the terms of the *) +(* * GNU Lesser General Public License Version 2.1 *) +(************************************************************************) + +(*i $Id: mlutil.ml,v 1.104.2.1 2004/07/16 19:30:08 herbelin Exp $ i*) + +(*i*) +open Pp +open Util +open Names +open Libnames +open Nametab +open Table +open Miniml +(*i*) + +(*s Exceptions. *) + +exception Found +exception Impossible + +(*S Names operations. *) + +let anonymous = id_of_string "x" +let dummy_name = id_of_string "_" + +let id_of_name = function + | Anonymous -> anonymous + | Name id when id = dummy_name -> anonymous + | Name id -> id + +(*S Operations upon ML types (with meta). *) + +let meta_count = ref 0 + +let reset_meta_count () = meta_count := 0 + +let new_meta _ = + incr meta_count; + Tmeta {id = !meta_count; contents = None} + +(*s Sustitution of [Tvar i] by [t] in a ML type. *) + +let type_subst i t0 t = + let rec subst t = match t with + | Tvar j when i = j -> t0 + | Tmeta {contents=None} -> t + | Tmeta {contents=Some u} -> subst u + | Tarr (a,b) -> Tarr (subst a, subst b) + | Tglob (r, l) -> Tglob (r, List.map subst l) + | a -> a + in subst t + +(* Simultaneous substitution of [[Tvar 1; ... ; Tvar n]] by [l] in a ML type. *) + +let type_subst_list l t = + let rec subst t = match t with + | Tvar j -> List.nth l (j-1) + | Tmeta {contents=None} -> t + | Tmeta {contents=Some u} -> subst u + | Tarr (a,b) -> Tarr (subst a, subst b) + | Tglob (r, l) -> Tglob (r, List.map subst l) + | a -> a + in subst t + +(* Simultaneous substitution of [[|Tvar 1; ... ; Tvar n|]] by [v] in a ML type. *) + +let type_subst_vect v t = + let rec subst t = match t with + | Tvar j -> v.(j-1) + | Tmeta {contents=None} -> t + | Tmeta {contents=Some u} -> subst u + | Tarr (a,b) -> Tarr (subst a, subst b) + | Tglob (r, l) -> Tglob (r, List.map subst l) + | a -> a + in subst t + +(*s From a type schema to a type. All [Tvar] become fresh [Tmeta]. *) + +let instantiation (nb,t) = type_subst_vect (Array.init nb new_meta) t + +(*s Occur-check of a free meta in a type *) + +let rec type_occurs alpha t = + match t with + | Tmeta {id=beta; contents=None} -> alpha = beta + | Tmeta {contents=Some u} -> type_occurs alpha u + | Tarr (t1, t2) -> type_occurs alpha t1 || type_occurs alpha t2 + | Tglob (r,l) -> List.exists (type_occurs alpha) l + | _ -> false + +(*s Most General Unificator *) + +let rec mgu = function + | Tmeta m, Tmeta m' when m.id = m'.id -> () + | Tmeta m, t when m.contents=None -> + if type_occurs m.id t then raise Impossible + else m.contents <- Some t + | t, Tmeta m when m.contents=None -> + if type_occurs m.id t then raise Impossible + else m.contents <- Some t + | Tmeta {contents=Some u}, t -> mgu (u, t) + | t, Tmeta {contents=Some u} -> mgu (t, u) + | Tarr(a, b), Tarr(a', b') -> + mgu (a, a'); mgu (b, b') + | Tglob (r,l), Tglob (r',l') when r = r' -> + List.iter mgu (List.combine l l') + | Tvar i, Tvar j when i = j -> () + | Tvar' i, Tvar' j when i = j -> () + | Tdummy, Tdummy -> () + | Tunknown, Tunknown -> () + | _ -> raise Impossible + +let needs_magic p = try mgu p; false with Impossible -> true + +let put_magic_if b a = if b then MLmagic a else a + +let put_magic p a = if needs_magic p then MLmagic a else a + + +(*S ML type env. *) + +module Mlenv = struct + + let meta_cmp m m' = compare m.id m'.id + module Metaset = Set.Make(struct type t = ml_meta let compare = meta_cmp end) + + (* Main MLenv type. [env] is the real environment, whereas [free] + (tries to) record the free meta variables occurring in [env]. *) + + type t = { env : ml_schema list; mutable free : Metaset.t} + + (* Empty environment. *) + + let empty = { env = []; free = Metaset.empty } + + (* [get] returns a instantiated copy of the n-th most recently added + type in the environment. *) + + let get mle n = + assert (List.length mle.env >= n); + instantiation (List.nth mle.env (n-1)) + + (* [find_free] finds the free meta in a type. *) + + let rec find_free set = function + | Tmeta m when m.contents = None -> Metaset.add m set + | Tmeta {contents = Some t} -> find_free set t + | Tarr (a,b) -> find_free (find_free set a) b + | Tglob (_,l) -> List.fold_left find_free set l + | _ -> set + + (* The [free] set of an environment can be outdate after + some unifications. [clean_free] takes care of that. *) + + let clean_free mle = + let rem = ref Metaset.empty + and add = ref Metaset.empty in + let clean m = match m.contents with + | None -> () + | Some u -> rem := Metaset.add m !rem; add := find_free !add u + in + Metaset.iter clean mle.free; + mle.free <- Metaset.union (Metaset.diff mle.free !rem) !add + + (* From a type to a type schema. If a [Tmeta] is still uninstantiated + and does appears in the [mle], then it becomes a [Tvar]. *) + + let generalization mle t = + let c = ref 0 in + let map = ref (Intmap.empty : int Intmap.t) in + let add_new i = incr c; map := Intmap.add i !c !map; !c in + let rec meta2var t = match t with + | Tmeta {contents=Some u} -> meta2var u + | Tmeta ({id=i} as m) -> + (try Tvar (Intmap.find i !map) + with Not_found -> + if Metaset.mem m mle.free then t + else Tvar (add_new i)) + | Tarr (t1,t2) -> Tarr (meta2var t1, meta2var t2) + | Tglob (r,l) -> Tglob (r, List.map meta2var l) + | t -> t + in !c, meta2var t + + (* Adding a type in an environment, after generalizing. *) + + let push_gen mle t = + clean_free mle; + { env = generalization mle t :: mle.env; free = mle.free } + + (* Adding a type with no [Tvar], hence no generalization needed. *) + + let push_type {env=e;free=f} t = + { env = (0,t) :: e; free = find_free f t} + + (* Adding a type with no [Tvar] nor [Tmeta]. *) + + let push_std_type {env=e;free=f} t = + { env = (0,t) :: e; free = f} + +end + +(*S Operations upon ML types (without meta). *) + +(*s Does a section path occur in a ML type ? *) + +let rec type_mem_kn kn = function + | Tmeta _ -> assert false + | Tglob (r,l) -> (kn_of_r r) = kn || List.exists (type_mem_kn kn) l + | Tarr (a,b) -> (type_mem_kn kn a) || (type_mem_kn kn b) + | _ -> false + +(*s Greatest variable occurring in [t]. *) + +let type_maxvar t = + let rec parse n = function + | Tmeta _ -> assert false + | Tvar i -> max i n + | Tarr (a,b) -> parse (parse n a) b + | Tglob (_,l) -> List.fold_left parse n l + | _ -> n + in parse 0 t + +(*s From [a -> b -> c] to [[a;b],c]. *) + +let rec type_decomp = function + | Tmeta _ -> assert false + | Tarr (a,b) -> let l,h = type_decomp b in a::l, h + | a -> [],a + +(*s The converse: From [[a;b],c] to [a -> b -> c]. *) + +let rec type_recomp (l,t) = match l with + | [] -> t + | a::l -> Tarr (a, type_recomp (l,t)) + +(*s Translating [Tvar] to [Tvar'] to avoid clash. *) + +let rec var2var' = function + | Tmeta _ -> assert false + | Tvar i -> Tvar' i + | Tarr (a,b) -> Tarr (var2var' a, var2var' b) + | Tglob (r,l) -> Tglob (r, List.map var2var' l) + | a -> a + +type abbrev_map = global_reference -> ml_type option + +(*s Delta-reduction of type constants everywhere in a ML type [t]. + [env] is a function of type [ml_type_env]. *) + +let type_expand env t = + let rec expand = function + | Tmeta _ -> assert false + | Tglob (r,l) as t -> + (match env r with + | Some mlt -> expand (type_subst_list l mlt) + | None -> Tglob (r, List.map expand l)) + | Tarr (a,b) -> Tarr (expand a, expand b) + | a -> a + in expand t + +(*s Idem, but only at the top level of implications. *) + +let is_arrow = function Tarr _ -> true | _ -> false + +let type_weak_expand env t = + let rec expand = function + | Tmeta _ -> assert false + | Tglob (r,l) as t -> + (match env r with + | Some mlt -> + let u = expand (type_subst_list l mlt) in + if is_arrow u then u else t + | None -> t) + | Tarr (a,b) -> Tarr (a, expand b) + | a -> a + in expand t + +(*s Equality over ML types modulo delta-reduction *) + +let type_eq env t t' = (type_expand env t = type_expand env t') + +let type_neq env t t' = (type_expand env t <> type_expand env t') + +(*s Generating a signature from a ML type. *) + +let type_to_sign env t = + let rec f = function + | Tmeta _ -> assert false + | Tarr (a,b) -> (Tdummy <> a) :: (f b) + | _ -> [] + in f (type_expand env t) + +(*s Removing [Tdummy] from the top level of a ML type. *) + +let type_expunge env t = + let s = type_to_sign env t in + if s = [] then t + else if List.mem true s then + let rec f t s = + if List.mem false s then + match t with + | Tmeta _ -> assert false + | Tarr (a,b) -> + let t = f b (List.tl s) in + if List.hd s then Tarr (a, t) else t + | Tglob (r,l) -> + (match env r with + | Some mlt -> f (type_subst_list l mlt) s + | None -> assert false) + | _ -> assert false + else t + in f t s + else Tarr (Tdummy, snd (type_decomp (type_weak_expand env t))) + +(*S Generic functions over ML ast terms. *) + +(*s [ast_iter_rel f t] applies [f] on every [MLrel] in t. It takes care + of the number of bingings crossed before reaching the [MLrel]. *) + +let ast_iter_rel f = + let rec iter n = function + | MLrel i -> f (i-n) + | MLlam (_,a) -> iter (n+1) a + | MLletin (_,a,b) -> iter n a; iter (n+1) b + | MLcase (a,v) -> + iter n a; Array.iter (fun (_,l,t) -> iter (n + (List.length l)) t) v + | MLfix (_,ids,v) -> let k = Array.length ids in Array.iter (iter (n+k)) v + | MLapp (a,l) -> iter n a; List.iter (iter n) l + | MLcons (_,l) -> List.iter (iter n) l + | MLmagic a -> iter n a + | MLglob _ | MLexn _ | MLdummy | MLaxiom -> () + in iter 0 + +(*s Map over asts. *) + +let ast_map_case f (c,ids,a) = (c,ids,f a) + +let ast_map f = function + | MLlam (i,a) -> MLlam (i, f a) + | MLletin (i,a,b) -> MLletin (i, f a, f b) + | MLcase (a,v) -> MLcase (f a, Array.map (ast_map_case f) v) + | MLfix (i,ids,v) -> MLfix (i, ids, Array.map f v) + | MLapp (a,l) -> MLapp (f a, List.map f l) + | MLcons (c,l) -> MLcons (c, List.map f l) + | MLmagic a -> MLmagic (f a) + | MLrel _ | MLglob _ | MLexn _ | MLdummy | MLaxiom as a -> a + +(*s Map over asts, with binding depth as parameter. *) + +let ast_map_lift_case f n (c,ids,a) = (c,ids, f (n+(List.length ids)) a) + +let ast_map_lift f n = function + | MLlam (i,a) -> MLlam (i, f (n+1) a) + | MLletin (i,a,b) -> MLletin (i, f n a, f (n+1) b) + | MLcase (a,v) -> MLcase (f n a,Array.map (ast_map_lift_case f n) v) + | MLfix (i,ids,v) -> + let k = Array.length ids in MLfix (i,ids,Array.map (f (k+n)) v) + | MLapp (a,l) -> MLapp (f n a, List.map (f n) l) + | MLcons (c,l) -> MLcons (c, List.map (f n) l) + | MLmagic a -> MLmagic (f n a) + | MLrel _ | MLglob _ | MLexn _ | MLdummy | MLaxiom as a -> a + +(*s Iter over asts. *) + +let ast_iter_case f (c,ids,a) = f a + +let ast_iter f = function + | MLlam (i,a) -> f a + | MLletin (i,a,b) -> f a; f b + | MLcase (a,v) -> f a; Array.iter (ast_iter_case f) v + | MLfix (i,ids,v) -> Array.iter f v + | MLapp (a,l) -> f a; List.iter f l + | MLcons (c,l) -> List.iter f l + | MLmagic a -> f a + | MLrel _ | MLglob _ | MLexn _ | MLdummy | MLaxiom as a -> () + +(*S Operations concerning De Bruijn indices. *) + +(*s [ast_occurs k t] returns [true] if [(Rel k)] occurs in [t]. *) + +let ast_occurs k t = + try + ast_iter_rel (fun i -> if i = k then raise Found) t; false + with Found -> true + +(*s [occurs_itvl k k' t] returns [true] if there is a [(Rel i)] + in [t] with [k<=i<=k'] *) + +let ast_occurs_itvl k k' t = + try + ast_iter_rel (fun i -> if (k <= i) && (i <= k') then raise Found) t; false + with Found -> true + +(*s Number of occurences of [Rel k] and [Rel 1] in [t]. *) + +let nb_occur_k k t = + let cpt = ref 0 in + ast_iter_rel (fun i -> if i = k then incr cpt) t; + !cpt + +let nb_occur t = nb_occur_k 1 t + +(* Number of occurences of [Rel 1] in [t], with special treatment of match: + occurences in different branches aren't added, but we rather use max. *) + +let nb_occur_match = + let rec nb k = function + | MLrel i -> if i = k then 1 else 0 + | MLcase(a,v) -> + (nb k a) + + Array.fold_left + (fun r (_,ids,a) -> max r (nb (k+(List.length ids)) a)) 0 v + | MLletin (_,a,b) -> (nb k a) + (nb (k+1) b) + | MLfix (_,ids,v) -> let k = k+(Array.length ids) in + Array.fold_left (fun r a -> r+(nb k a)) 0 v + | MLlam (_,a) -> nb (k+1) a + | MLapp (a,l) -> List.fold_left (fun r a -> r+(nb k a)) (nb k a) l + | MLcons (_,l) -> List.fold_left (fun r a -> r+(nb k a)) 0 l + | MLmagic a -> nb k a + | MLglob _ | MLexn _ | MLdummy | MLaxiom -> 0 + in nb 1 + +(*s Lifting on terms. + [ast_lift k t] lifts the binding depth of [t] across [k] bindings. *) + +let ast_lift k t = + let rec liftrec n = function + | MLrel i as a -> if i-n < 1 then a else MLrel (i+k) + | a -> ast_map_lift liftrec n a + in if k = 0 then t else liftrec 0 t + +let ast_pop t = ast_lift (-1) t + +(*s [permut_rels k k' c] translates [Rel 1 ... Rel k] to [Rel (k'+1) ... + Rel (k'+k)] and [Rel (k+1) ... Rel (k+k')] to [Rel 1 ... Rel k'] *) + +let permut_rels k k' = + let rec permut n = function + | MLrel i as a -> + let i' = i-n in + if i'<1 || i'>k+k' then a + else if i'<=k then MLrel (i+k') + else MLrel (i-k) + | a -> ast_map_lift permut n a + in permut 0 + +(*s Substitution. [ml_subst e t] substitutes [e] for [Rel 1] in [t]. + Lifting (of one binder) is done at the same time. *) + +let ast_subst e = + let rec subst n = function + | MLrel i as a -> + let i' = i-n in + if i'=1 then ast_lift n e + else if i'<1 then a + else MLrel (i-1) + | a -> ast_map_lift subst n a + in subst 0 + +(*s Generalized substitution. + [gen_subst v d t] applies to [t] the substitution coded in the + [v] array: [(Rel i)] becomes [v.(i-1)]. [d] is the correction applies + to [Rel] greater than [Array.length v]. *) + +let gen_subst v d t = + let rec subst n = function + | MLrel i as a -> + let i'= i-n in + if i' < 1 then a + else if i' <= Array.length v then + ast_lift n v.(i'-1) + else MLrel (i+d) + | a -> ast_map_lift subst n a + in subst 0 t + +(*S Operations concerning lambdas. *) + +(*s [collect_lams MLlam(id1,...MLlam(idn,t)...)] returns + [[idn;...;id1]] and the term [t]. *) + +let collect_lams = + let rec collect acc = function + | MLlam(id,t) -> collect (id::acc) t + | x -> acc,x + in collect [] + +(*s [collect_n_lams] does the same for a precise number of [MLlam]. *) + +let collect_n_lams = + let rec collect acc n t = + if n = 0 then acc,t + else match t with + | MLlam(id,t) -> collect (id::acc) (n-1) t + | _ -> assert false + in collect [] + +(*s [remove_n_lams] just removes some [MLlam]. *) + +let rec remove_n_lams n t = + if n = 0 then t + else match t with + | MLlam(_,t) -> remove_n_lams (n-1) t + | _ -> assert false + +(*s [nb_lams] gives the number of head [MLlam]. *) + +let rec nb_lams = function + | MLlam(_,t) -> succ (nb_lams t) + | _ -> 0 + +(*s [named_lams] does the converse of [collect_lams]. *) + +let rec named_lams ids a = match ids with + | [] -> a + | id :: ids -> named_lams ids (MLlam (id,a)) + +(*s The same in anonymous version. *) + +let rec anonym_lams a = function + | 0 -> a + | n -> anonym_lams (MLlam (anonymous,a)) (pred n) + +(*s Idem for [dummy_name]. *) + +let rec dummy_lams a = function + | 0 -> a + | n -> dummy_lams (MLlam (dummy_name,a)) (pred n) + +(*s mixed according to a signature. *) + +let rec anonym_or_dummy_lams a = function + | [] -> a + | true :: s -> MLlam(anonymous, anonym_or_dummy_lams a s) + | false :: s -> MLlam(dummy_name, anonym_or_dummy_lams a s) + +(*S Operations concerning eta. *) + +(*s The following function creates [MLrel n;...;MLrel 1] *) + +let rec eta_args n = + if n = 0 then [] else (MLrel n)::(eta_args (pred n)) + +(*s Same, but filtered by a signature. *) + +let rec eta_args_sign n = function + | [] -> [] + | true :: s -> (MLrel n) :: (eta_args_sign (n-1) s) + | false :: s -> eta_args_sign (n-1) s + +(*s This one tests [MLrel (n+k); ... ;MLrel (1+k)] *) + +let rec test_eta_args_lift k n = function + | [] -> n=0 + | a :: q -> (a = (MLrel (k+n))) && (test_eta_args_lift k (pred n) q) + +(*s Computes an eta-reduction. *) + +let eta_red e = + let ids,t = collect_lams e in + let n = List.length ids in + if n = 0 then e + else match t with + | MLapp (f,a) -> + let m = (List.length a) - n in + if m < 0 then e + else + let a1,a2 = list_chop m a in + let f = if m = 0 then f else MLapp (f,a1) in + if test_eta_args_lift 0 n a2 && not (ast_occurs_itvl 1 n f) + then ast_lift (-n) f + else e + | _ -> e + +(*s Computes all head linear beta-reductions possible in [(t a)]. + Non-linear head beta-redex become let-in. *) + +let rec linear_beta_red a t = match a,t with + | [], _ -> t + | a0::a, MLlam (id,t) -> + (match nb_occur_match t with + | 0 -> linear_beta_red a (ast_pop t) + | 1 -> linear_beta_red a (ast_subst a0 t) + | _ -> + let a = List.map (ast_lift 1) a in + MLletin (id, a0, linear_beta_red a t)) + | _ -> MLapp (t, a) + +(*s Applies a substitution [s] of constants by their body, plus + linear beta reductions at modified positions. *) + +let rec ast_glob_subst s t = match t with + | MLapp ((MLglob (ConstRef kn)) as f, a) -> + let a = List.map (ast_glob_subst s) a in + (try linear_beta_red a (KNmap.find kn s) + with Not_found -> MLapp (f, a)) + | MLglob (ConstRef kn) -> (try KNmap.find kn s with Not_found -> t) + | _ -> ast_map (ast_glob_subst s) t + + +(*S Auxiliary functions used in simplification of ML cases. *) + +(*s [check_and_generalize (r0,l,c)] transforms any [MLcons(r0,l)] in [MLrel 1] + and raises [Impossible] if any variable in [l] occurs outside such a + [MLcons] *) + +let check_and_generalize (r0,l,c) = + let nargs = List.length l in + let rec genrec n = function + | MLrel i as c -> + let i' = i-n in + if i'<1 then c + else if i'>nargs then MLrel (i-nargs+1) + else raise Impossible + | MLcons(r,args) when r=r0 && (test_eta_args_lift n nargs args) -> + MLrel (n+1) + | a -> ast_map_lift genrec n a + in genrec 0 c + +(*s [check_generalizable_case] checks if all branches can be seen as the + same function [f] applied to the term matched. It is a generalized version + of the identity case optimization. *) + +(* CAVEAT: this optimization breaks typing in some special case. example: + [type 'x a = A]. Then [let f = function A -> A] has type ['x a -> 'y a], + which is incompatible with the type of [let f x = x]. + By default, we brutally disable this optim except for some known types: + [bool], [sumbool], [sumor] *) + +let generalizable_list = + let datatypes = MPfile (dirpath_of_string "Coq.Init.Datatypes") + and specif = MPfile (dirpath_of_string "Coq.Init.Specif") + in + [ make_kn datatypes empty_dirpath (mk_label "bool"); + make_kn specif empty_dirpath (mk_label "sumbool"); + make_kn specif empty_dirpath (mk_label "sumor") ] + +let check_generalizable_case unsafe br = + if not unsafe then + (match br.(0) with + | ConstructRef ((kn,_),_), _, _ -> + if not (List.mem kn generalizable_list) then raise Impossible + | _ -> assert false); + let f = check_and_generalize br.(0) in + for i = 1 to Array.length br - 1 do + if check_and_generalize br.(i) <> f then raise Impossible + done; f + +(*s Do all branches correspond to the same thing? *) + +let check_constant_case br = + if br = [||] then raise Impossible; + let (r,l,t) = br.(0) in + let n = List.length l in + if ast_occurs_itvl 1 n t then raise Impossible; + let cst = ast_lift (-n) t in + for i = 1 to Array.length br - 1 do + let (r,l,t) = br.(i) in + let n = List.length l in + if (ast_occurs_itvl 1 n t) || (cst <> (ast_lift (-n) t)) + then raise Impossible + done; cst + +(*s If all branches are functions, try to permut the case and the functions. *) + +let rec merge_ids ids ids' = match ids,ids' with + | [],l -> l + | l,[] -> l + | i::ids, i'::ids' -> + (if i = dummy_name then i' else i) :: (merge_ids ids ids') + +let is_exn = function MLexn _ -> true | _ -> false + +let rec permut_case_fun br acc = + let nb = ref max_int in + Array.iter (fun (_,_,t) -> + let ids, c = collect_lams t in + let n = List.length ids in + if (n < !nb) && (not (is_exn c)) then nb := n) br; + if !nb = max_int || !nb = 0 then ([],br) + else begin + let br = Array.copy br in + let ids = ref [] in + for i = 0 to Array.length br - 1 do + let (r,l,t) = br.(i) in + let local_nb = nb_lams t in + if local_nb < !nb then (* t = MLexn ... *) + br.(i) <- (r,l,remove_n_lams local_nb t) + else begin + let local_ids,t = collect_n_lams !nb t in + ids := merge_ids !ids local_ids; + br.(i) <- (r,l,permut_rels !nb (List.length l) t) + end + done; + (!ids,br) + end + +(*S Generalized iota-reduction. *) + +(* Definition of a generalized iota-redex: it's a [MLcase(e,_)] + with [(is_iota_gen e)=true]. Any generalized iota-redex is + transformed into beta-redexes. *) + +let rec is_iota_gen = function + | MLcons _ -> true + | MLcase(_,br)-> array_for_all (fun (_,_,t)->is_iota_gen t) br + | _ -> false + +let constructor_index = function + | ConstructRef (_,j) -> pred j + | _ -> assert false + +let iota_gen br = + let rec iota k = function + | MLcons (r,a) -> + let (_,ids,c) = br.(constructor_index r) in + let c = List.fold_right (fun id t -> MLlam (id,t)) ids c in + let c = ast_lift k c in + MLapp (c,a) + | MLcase(e,br') -> + let new_br = + Array.map (fun (n,i,c)->(n,i,iota (k+(List.length i)) c)) br' + in MLcase(e, new_br) + | _ -> assert false + in iota 0 + +let is_atomic = function + | MLrel _ | MLglob _ | MLexn _ | MLdummy -> true + | _ -> false + +(*S The main simplification function. *) + +(* Some beta-iota reductions + simplifications. *) + +let rec simpl o = function + | MLapp (f, []) -> + simpl o f + | MLapp (f, a) -> + simpl_app o (List.map (simpl o) a) (simpl o f) + | MLcase (e,br) -> + let br = Array.map (fun (n,l,t) -> (n,l,simpl o t)) br in + simpl_case o br (simpl o e) + | MLletin(id,c,e) when + (id = dummy_name) || (is_atomic c) || (is_atomic e) || + (let n = nb_occur_match e in n = 0 || (n=1 && o.opt_lin_let)) -> + simpl o (ast_subst c e) + | MLfix(i,ids,c) -> + let n = Array.length ids in + if ast_occurs_itvl 1 n c.(i) then + MLfix (i, ids, Array.map (simpl o) c) + else simpl o (ast_lift (-n) c.(i)) (* Dummy fixpoint *) + | a -> ast_map (simpl o) a + +and simpl_app o a = function + | MLapp (f',a') -> simpl_app o (a'@a) f' + | MLlam (id,t) when id = dummy_name -> + simpl o (MLapp (ast_pop t, List.tl a)) + | MLlam (id,t) -> (* Beta redex *) + (match nb_occur_match t with + | 0 -> simpl o (MLapp (ast_pop t, List.tl a)) + | 1 when o.opt_lin_beta -> + simpl o (MLapp (ast_subst (List.hd a) t, List.tl a)) + | _ -> + let a' = List.map (ast_lift 1) (List.tl a) in + simpl o (MLletin (id, List.hd a, MLapp (t, a')))) + | MLletin (id,e1,e2) when o.opt_let_app -> + (* Application of a letin: we push arguments inside *) + MLletin (id, e1, simpl o (MLapp (e2, List.map (ast_lift 1) a))) + | MLcase (e,br) when o.opt_case_app -> + (* Application of a case: we push arguments inside *) + let br' = + Array.map + (fun (n,l,t) -> + let k = List.length l in + let a' = List.map (ast_lift k) a in + (n, l, simpl o (MLapp (t,a')))) br + in simpl o (MLcase (e,br')) + | (MLdummy | MLexn _) as e -> e + (* We just discard arguments in those cases. *) + | f -> MLapp (f,a) + +and simpl_case o br e = + if o.opt_case_iot && (is_iota_gen e) then (* Generalized iota-redex *) + simpl o (iota_gen br e) + else + try (* Does a term [f] exist such as each branch is [(f e)] ? *) + if not o.opt_case_idr then raise Impossible; + let f = check_generalizable_case o.opt_case_idg br in + simpl o (MLapp (MLlam (anonymous,f),[e])) + with Impossible -> + try (* Is each branch independant of [e] ? *) + if not o.opt_case_cst then raise Impossible; + check_constant_case br + with Impossible -> + (* Swap the case and the lam if possible *) + if o.opt_case_fun + then + let ids,br = permut_case_fun br [] in + let n = List.length ids in + if n <> 0 then named_lams ids (MLcase (ast_lift n e, br)) + else MLcase (e, br) + else MLcase (e,br) + +let rec post_simpl = function + | MLletin(_,c,e) when (is_atomic (eta_red c)) -> + post_simpl (ast_subst (eta_red c) e) + | a -> ast_map post_simpl a + +(*S Local prop elimination. *) +(* We try to eliminate as many [prop] as possible inside an [ml_ast]. *) + +(*s In a list, it selects only the elements corresponding to a [true] + in the boolean list [l]. *) + +let rec select_via_bl l args = match l,args with + | [],_ -> args + | true::l,a::args -> a :: (select_via_bl l args) + | false::l,a::args -> select_via_bl l args + | _ -> assert false + +(*s [kill_some_lams] removes some head lambdas according to the bool list [bl]. + This list is build on the identifier list model: outermost lambda + is on the right. [true] means "to keep" and [false] means "to eliminate". + [Rels] corresponding to removed lambdas are supposed not to occur, and + the other [Rels] are made correct via a [gen_subst]. + Output is not directly a [ml_ast], compose with [named_lams] if needed. *) + +let kill_some_lams bl (ids,c) = + let n = List.length bl in + let n' = List.fold_left (fun n b -> if b then (n+1) else n) 0 bl in + if n = n' then ids,c + else if n' = 0 then [],ast_lift (-n) c + else begin + let v = Array.make n MLdummy in + let rec parse_ids i j = function + | [] -> () + | true :: l -> v.(i) <- MLrel j; parse_ids (i+1) (j+1) l + | false :: l -> parse_ids (i+1) j l + in parse_ids 0 1 bl ; + select_via_bl bl ids, gen_subst v (n'-n) c + end + +(*s [kill_dummy_lams] uses the last function to kill the lambdas corresponding + to a [dummy_name]. It can raise [Impossible] if there is nothing to do, or + if there is no lambda left at all. *) + +let kill_dummy_lams c = + let ids,c = collect_lams c in + let bl = List.map ((<>) dummy_name) ids in + if (List.mem true bl) && (List.mem false bl) then + let ids',c = kill_some_lams bl (ids,c) in + ids, named_lams ids' c + else raise Impossible + +(*s [eta_expansion_sign] takes a function [fun idn ... id1 -> c] + and a signature [s] and builds a eta-long version. *) + +(* For example, if [s = [true;true;false;true]] then the output is : + [fun idn ... id1 x x _ x -> (c' 4 3 __ 1)] with [c' = lift 4 c] *) + +let eta_expansion_sign s (ids,c) = + let rec abs ids rels i = function + | [] -> + let a = List.rev_map (function MLrel x -> MLrel (i-x) | a -> a) rels + in ids, MLapp (ast_lift (i-1) c, a) + | true :: l -> abs (anonymous :: ids) (MLrel i :: rels) (i+1) l + | false :: l -> abs (dummy_name :: ids) (MLdummy :: rels) (i+1) l + in abs ids [] 1 s + +(*s If [s = [b1; ... ; bn]] then [case_expunge] decomposes [e] + in [n] lambdas (with eta-expansion if needed) and removes all dummy lambdas + corresponding to [false] in [s]. *) + +let case_expunge s e = + let m = List.length s in + let n = nb_lams e in + let p = if m <= n then collect_n_lams m e + else eta_expansion_sign (list_skipn n s) (collect_lams e) in + kill_some_lams (List.rev s) p + +(*s [term_expunge] takes a function [fun idn ... id1 -> c] + and a signature [s] and remove dummy lams. The difference + with [case_expunge] is that we here leave one dummy lambda + if all lambdas are dummy. *) + +let term_expunge s (ids,c) = + if s = [] then c + else + let ids,c = kill_some_lams (List.rev s) (ids,c) in + if ids = [] then MLlam (dummy_name, ast_lift 1 c) + else named_lams ids c + +(*s [kill_dummy_args ids t0 t] looks for occurences of [t0] in [t] and + purge the args of [t0] corresponding to a [dummy_name]. + It makes eta-expansion if needed. *) + +let kill_dummy_args ids t0 t = + let m = List.length ids in + let bl = List.rev_map ((<>) dummy_name) ids in + let rec killrec n = function + | MLapp(e, a) when e = ast_lift n t0 -> + let k = max 0 (m - (List.length a)) in + let a = List.map (killrec n) a in + let a = List.map (ast_lift k) a in + let a = select_via_bl bl (a @ (eta_args k)) in + named_lams (list_firstn k ids) (MLapp (ast_lift k e, a)) + | e when e = ast_lift n t0 -> + let a = select_via_bl bl (eta_args m) in + named_lams ids (MLapp (ast_lift m e, a)) + | e -> ast_map_lift killrec n e + in killrec 0 t + +(*s The main function for local [dummy] elimination. *) + +let rec kill_dummy = function + | MLfix(i,fi,c) -> + (try + let ids,c = kill_dummy_fix i fi c in + ast_subst (MLfix (i,fi,c)) (kill_dummy_args ids (MLrel 1) (MLrel 1)) + with Impossible -> MLfix (i,fi,Array.map kill_dummy c)) + | MLapp (MLfix (i,fi,c),a) -> + (try + let ids,c = kill_dummy_fix i fi c in + let a = List.map (fun t -> ast_lift 1 (kill_dummy t)) a in + let e = kill_dummy_args ids (MLrel 1) (MLapp (MLrel 1,a)) in + ast_subst (MLfix (i,fi,c)) e + with Impossible -> + MLapp(MLfix(i,fi,Array.map kill_dummy c),List.map kill_dummy a)) + | MLletin(id, MLfix (i,fi,c),e) -> + (try + let ids,c = kill_dummy_fix i fi c in + let e = kill_dummy (kill_dummy_args ids (MLrel 1) e) in + MLletin(id, MLfix(i,fi,c),e) + with Impossible -> + MLletin(id, MLfix(i,fi,Array.map kill_dummy c),kill_dummy e)) + | MLletin(id,c,e) -> + (try + let ids,c = kill_dummy_lams c in + let e = kill_dummy_args ids (MLrel 1) e in + MLletin (id, kill_dummy c,kill_dummy e) + with Impossible -> MLletin(id,kill_dummy c,kill_dummy e)) + | a -> ast_map kill_dummy a + +and kill_dummy_fix i fi c = + let n = Array.length fi in + let ids,ci = kill_dummy_lams c.(i) in + let c = Array.copy c in c.(i) <- ci; + for j = 0 to (n-1) do + c.(j) <- kill_dummy (kill_dummy_args ids (MLrel (n-i)) c.(j)) + done; + ids,c + +(*s Putting things together. *) + +let normalize a = + let o = optims () in + let a = simpl o a in + if o.opt_kill_dum then post_simpl (kill_dummy a) else a + +(*S Special treatment of fixpoint for pretty-printing purpose. *) + +let general_optimize_fix f ids n args m c = + let v = Array.make n 0 in + for i=0 to (n-1) do v.(i)<-i done; + let aux i = function + | MLrel j when v.(j-1)>=0 -> v.(j-1)<-(-i-1) + | _ -> raise Impossible + in list_iter_i aux args; + let args_f = List.rev_map (fun i -> MLrel (i+m+1)) (Array.to_list v) in + let new_f = anonym_lams (MLapp (MLrel (n+m+1),args_f)) m in + let new_c = named_lams ids (normalize (MLapp ((ast_subst new_f c),args))) in + MLfix(0,[|f|],[|new_c|]) + +let optimize_fix a = + if not (optims()).opt_fix_fun then a + else + let ids,a' = collect_lams a in + let n = List.length ids in + if n = 0 then a + else match a' with + | MLfix(_,[|f|],[|c|]) -> + let new_f = MLapp (MLrel (n+1),eta_args n) in + let new_c = named_lams ids (normalize (ast_subst new_f c)) + in MLfix(0,[|f|],[|new_c|]) + | MLapp(a',args) -> + let m = List.length args in + (match a' with + | MLfix(_,_,_) when + (test_eta_args_lift 0 n args) && not (ast_occurs_itvl 1 m a') + -> a' + | MLfix(_,[|f|],[|c|]) -> + (try general_optimize_fix f ids n args m c + with Impossible -> + named_lams ids (MLapp (MLfix (0,[|f|],[|c|]),args))) + | _ -> a) + | _ -> a + +(*S Inlining. *) + +(* Utility functions used in the decision of inlining. *) + +let rec ml_size = function + | MLapp(t,l) -> List.length l + ml_size t + ml_size_list l + | MLlam(_,t) -> 1 + ml_size t + | MLcons(_,l) -> ml_size_list l + | MLcase(t,pv) -> + 1 + ml_size t + (Array.fold_right (fun (_,_,t) a -> a + ml_size t) pv 0) + | MLfix(_,_,f) -> ml_size_array f + | MLletin (_,_,t) -> ml_size t + | MLmagic t -> ml_size t + | _ -> 0 + +and ml_size_list l = List.fold_left (fun a t -> a + ml_size t) 0 l + +and ml_size_array l = Array.fold_left (fun a t -> a + ml_size t) 0 l + +let is_fix = function MLfix _ -> true | _ -> false + +let rec is_constr = function + | MLcons _ -> true + | MLlam(_,t) -> is_constr t + | _ -> false + +(*s Strictness *) + +(* A variable is strict if the evaluation of the whole term implies + the evaluation of this variable. Non-strict variables can be found + behind Match, for example. Expanding a term [t] is a good idea when + it begins by at least one non-strict lambda, since the corresponding + argument to [t] might be unevaluated in the expanded code. *) + +exception Toplevel + +let lift n l = List.map ((+) n) l + +let pop n l = List.map (fun x -> if x<=n then raise Toplevel else x-n) l + +(* This function returns a list of de Bruijn indices of non-strict variables, + or raises [Toplevel] if it has an internal non-strict variable. + In fact, not all variables are checked for strictness, only the ones which + de Bruijn index is in the candidates list [cand]. The flag [add] controls + the behaviour when going through a lambda: should we add the corresponding + variable to the candidates? We use this flag to check only the external + lambdas, those that will correspond to arguments. *) + +let rec non_stricts add cand = function + | MLlam (id,t) -> + let cand = lift 1 cand in + let cand = if add then 1::cand else cand in + pop 1 (non_stricts add cand t) + | MLrel n -> + List.filter ((<>) n) cand + | MLapp (MLrel n, _) -> + List.filter ((<>) n) cand + (* In [(x y)] we say that only x is strict. Cf [sig_rec]. We may *) + (* gain something if x is replaced by a function like a projection *) + | MLapp (t,l)-> + let cand = non_stricts false cand t in + List.fold_left (non_stricts false) cand l + | MLcons (_,l) -> + List.fold_left (non_stricts false) cand l + | MLletin (_,t1,t2) -> + let cand = non_stricts false cand t1 in + pop 1 (non_stricts add (lift 1 cand) t2) + | MLfix (_,i,f)-> + let n = Array.length i in + let cand = lift n cand in + let cand = Array.fold_left (non_stricts false) cand f in + pop n cand + | MLcase (t,v) -> + (* The only interesting case: for a variable to be non-strict, *) + (* it is sufficient that it appears non-strict in at least one branch, *) + (* so we make an union (in fact a merge). *) + let cand = non_stricts false cand t in + Array.fold_left + (fun c (_,i,t)-> + let n = List.length i in + let cand = lift n cand in + let cand = pop n (non_stricts add cand t) in + Sort.merge (<=) cand c) [] v + (* [merge] may duplicates some indices, but I don't mind. *) + | MLmagic t -> + non_stricts add cand t + | _ -> + cand + +(* The real test: we are looking for internal non-strict variables, so we start + with no candidates, and the only positive answer is via the [Toplevel] + exception. *) + +let is_not_strict t = + try let _ = non_stricts true [] t in false + with Toplevel -> true + +(*s Inlining decision *) + +(* [inline_test] answers the following question: + If we could inline [t] (the user said nothing special), + should we inline ? + + We expand small terms with at least one non-strict + variable (i.e. a variable that may not be evaluated). + + Futhermore we don't expand fixpoints. *) + +let inline_test t = + not (is_fix (eta_red t)) && (ml_size t < 12 && is_not_strict t) + +let manual_inline_list = + let mp = MPfile (dirpath_of_string "Coq.Init.Wf") in + List.map (fun s -> (make_kn mp empty_dirpath (mk_label s))) + [ "well_founded_induction_type"; "well_founded_induction"; + "Acc_rect"; "Acc_rec" ; "Acc_iter" ] + +let manual_inline = function + | ConstRef c -> List.mem c manual_inline_list + | _ -> false + +(* If the user doesn't say he wants to keep [t], we inline in two cases: + \begin{itemize} + \item the user explicitly requests it + \item [expansion_test] answers that the inlining is a good idea, and + we are free to act (AutoInline is set) + \end{itemize} *) + +let inline r t = + not (to_keep r) (* The user DOES want to keep it *) + && not (is_inline_custom r) + && (to_inline r (* The user DOES want to inline it *) + || (auto_inline () && lang () <> Haskell && not (is_projection r) + && (is_recursor r || manual_inline r || inline_test t))) + |