path: root/compiler/Config.ml

(** Define the global configuration options *)

(** {1 Backend choice} *)

(** The choice of backend *)
type backend = FStar | Coq | Lean | HOL4

let backend_names = [ "fstar"; "coq"; "lean"; "hol4" ]

(** Utility to compute the backend from an input parameter *)
let backend_of_string (b : string) : backend option =
  match b with
  | "fstar" -> Some FStar
  | "coq" -> Some Coq
  | "lean" -> Some Lean
  | "hol4" -> Some HOL4
  | _ -> None

let opt_backend : backend option ref = ref None

let set_backend (b : string) : unit =
  match backend_of_string b with
  | Some b -> opt_backend := Some b
  | None ->
      (* We shouldn't get there: the string should have been checked as
         belonging to the proper set *)
      raise (Failure "Unexpected")

(** The backend to which to extract.

    We initialize it with a default value, but it always gets overwritten:
    we check that the user provides a backend argument.
 *)
let backend = ref FStar

(** {1 Interpreter} *)

(** Activate the sanity checks, and in particular the invariant checks
    that are performed at every evaluation step. This is very expensive
    (~100x slow down) but very efficient to catch mistakes early.
 *)
let sanity_checks = ref false

(** Expand all symbolic values containing borrows upon introduction - allows
    to use restrict ourselves to a simpler model for the projectors over
    symbolic values.
    The interpreter fails if doing this requires to do a branching (because
    we need to expand an enumeration with strictly more than one variant)
    or if we need to expand a recursive type (because this leads to looping).
 *)
let greedy_expand_symbolics_with_borrows = true

(** Experimental.

    TODO: remove (always true now), but check that when we panic/call a function
    there is no bottom below a borrow.

    We sometimes want to temporarily break the invariant that there is no
    bottom value below a borrow. If this value is true, we don't check
    the invariant, and the rule becomes: we can't end a borrow *if* it contains
    a bottom value. The consequence is that it becomes ok to temporarily
    have bottom below a borrow, if we put something else inside before ending
    the borrow.

    For instance, when evaluating an assignment, we move the value which
    will be overwritten then do some administrative tasks with the borrows,
    then move the rvalue to its destination. We currently want to be able
    to check the invariants every time we end a borrow/an abstraction,
    meaning at intermediate steps of the assignment where the invariants
    might actually be broken.
 *)
let allow_bottom_below_borrow = true

(** If a function doesn't return any borrows, we can immediately call
    its backward functions. If this option is on, whenever we call a
    function *and* this function returns unit, we immediately end all the
    abstractions which are introduced and don't contain loans. This can be
    useful to make the code cleaner (the backward function is introduced
    where the function call happened) and make sure all forward functions
    with no return value are followed by a backward function.
 *)
let return_unit_end_abs_with_no_loans = true

(** The maximum number of iterations we can do to find a loop fixed point.

    This is mostly for sanity: we should always find a fixed point in a
    reasonable number of iterations. If we fail to do so, it is likely a bug: we
    thus use this bound to detect a loop, fail and report the case to the
    user.
 *)
let loop_fixed_point_max_num_iters = 2

(** {1 Translation} *)

(** If true, do not define separate forward/backward functions, but make the
    forward functions return the backward function.

    Example:
    {[
      (* Rust *)
      pub fn list_nth<'a, T>(l: &'a mut List<T>, i: u32) -> &'a mut T {
          match l {
              List::Nil => {
                  panic!()
              }
              List::Cons(x, tl) => {
                  if i == 0 {
                      x
                  } else {
                      list_nth(tl, i - 1)
                  }
              }
          }
      }

      (* Translation, if return_back_funs = false *)
      def list_nth (T : Type) (l : List T) (i : U32) : Result T :=
        match l with
        | List.Cons x tl =>
          if i = 0#u32
          then Result.ret x
          else do
                 let i0 ← i - 1#u32
                 list_nth T tl i0
        | List.Nil => Result.fail .panic

      def list_nth_back
        (T : Type) (l : List T) (i : U32) (ret : T) : Result (List T) :=
        match l with
        | List.Cons x tl =>
          if i = 0#u32
          then Result.ret (List.Cons ret tl)
          else
            do
              let i0 ← i - 1#u32
              let tl0 ← list_nth_back T tl i0 ret
              Result.ret (List.Cons x tl0)
        | List.Nil => Result.fail .panic

      (* Translation, if return_back_funs = true *)
      def list_nth (T: Type) (ls : List T) (i : U32) :
        Result (T × (T → Result (List T))) :=
        match ls with
        | List.Cons x tl =>
          if i = 0#u32
          then Result.ret (x, (λ ret => return (ret :: ls)))
          else do
            let i0 ← i - 1#u32
            let (x, back) ← list_nth ls i0
            Return.ret (x,
              (λ ret => do
               let ls ← back ret
               return (x :: ls)))
    ]}
 *)
let return_back_funs = ref true

(** Forbids using field projectors for structures.

    If we don't use field projectors, whenever we symbolically expand a structure
    value (note that accessing a structure field in the symbolic execution triggers
    its expansion), then instead of generating code like this:
    {[
       let x1 = s.f1 in
       let x2 = s.f2 in
       ...
    ]}

    we generate code like this:
    {[
       let Mkstruct x1 x2 ... = s in
       ...
    ]}

    Rem.: this used to be useful for Coq, because in Coq we can't define groups
    of mutually recursive records and inductives. In such cases, we extract
    the structures as inductives, which means that field projectors are not
    always available. As of today, we generate field projectors definitions
    (together with the appropriate notations).
 *)
let dont_use_field_projectors = ref false

(** Deconstructing ADTs which have only one variant with let-bindings is not always
    supported: this parameter controls whether we use let-bindings in such situations or not.
  *)
let always_deconstruct_adts_with_matches = ref false

(** Controls whether we need to use a state to model the external world
    (I/O, for instance).
 *)
let use_state = ref false

(** Controls whether we use fuel to control termination.
 *)
let use_fuel = ref false

(** Controls whether backward functions update the state, in case we use
    a state ({!use_state}).

    If they update the state, we generate code of the following style:
    {[
       (st1, y)  <-- f_fwd x st0; // st0 is the state upon calling f_fwd
       ...
       (st3, x') <-- f_back x st0 y' st2; // st2 is the state upon calling f_back
    ]}

    Otherwise, we generate code of the following shape:
    {[
       (st1, y)  <-- f_fwd x st0;
       ...
       x' <-- f_back x st0 y';
    ]}

    The second format is easier to reason about, but the first one is
    necessary to properly handle some Rust functions which use internal
    mutability such as
    {{:https://doc.rust-lang.org/std/cell/struct.RefCell.html#method.try_borrow_mut}[RefCell::try_mut_borrow]}:
    in order to model this behaviour we would need a state, and calling the backward
    function would update the state by reinserting the updated value in it.
 *)
let backward_no_state_update = ref false

(** Controls whether we split the generated definitions between different
    files for the types, clauses and functions, or if we group them in
    one file.
 *)
let split_files = ref false

(** Generate the library entry point, if the crate is split between different files.

    Sometimes we want to skip this: the library entry points just includes all the
    files in the project, and the user may want to write their own entry point, to
    add custom includes (to include the files containing the proofs, for instance).
 *)
let generate_lib_entry_point = ref true

(** For Lean, controls whether we generate a lakefile or not.

 *)
let lean_gen_lakefile = ref false

(** If true, treat the unit functions (function taking no inputs and returning
    no outputs) as unit tests: evaluate them with the interpreter and check that
    they don't panic.
 *)
let test_unit_functions = ref false

(** If true, insert tests in the generated files to check that the
    unit functions normalize to [Success _].

    For instance, in F* it generates code like this:
    {[
      let _ = assert_norm (FUNCTION () = Success ())
    ]}
 *)
let test_trans_unit_functions = ref false

(** If [true], use decreases clauses/termination measures for all the recursive definitions.

    More precisely:
    - for F*, we generate definitions which use decreases clauses
    - for Lean, we generate definitions which use termination measures and
      decreases proofs

    The body of such clauses/proofs must be defined by the user.
 *)
let extract_decreases_clauses = ref false

(** In order to help the user, we can generate "template" decrease clauses/ termination
   measures (i.e., definitions with proper signatures but dummy bodies) in a dedicated
   file.
*)
let extract_template_decreases_clauses = ref false

(** {1 Micro passes} *)

(** Some provers like F* and Coq don't support the decomposition of return values
    in monadic let-bindings:
    {[
      (* NOT supported in F*/Coq *)
      (x, y) <-- f ();
      ...
    ]}

    In such situations, we might want to introduce an intermediate
    assignment:
    {[
      tmp <-- f ();
      let (x, y) = tmp in
      ...
    ]}
 *)
let decompose_monadic_let_bindings = ref false

(** Some provers like Coq don't support nested patterns in let-bindings:
    {[
      (* NOT supported in Coq *)
      (st, (x1, x2)) <-- f ();
      ...
    ]}

    In such situations, we might want to introduce intermediate
    assignments:
    {[
      (st, tmp) <-- f ();
      let (x1, x2) = tmp in
      ...
    ]}
 *)
let decompose_nested_let_patterns = ref false

(** Controls the unfolding of monadic let-bindings to explicit matches:

    [y <-- f x; ...]

    becomes:

    [match f x with | Failure -> Failure | Return y -> ...]


    This is useful when extracting to F*: the support for monadic
    definitions is not super powerful.
    Note that when {!unfold_monadic_let_bindings} is true, setting
    {!decompose_monadic_let_bindings} to true and only makes the code
    more verbose.
 *)
let unfold_monadic_let_bindings = ref false

(** Controls whether we try to filter the calls to monadic functions
    (which can fail) when their outputs are not used.

    The useless calls are calls to backward functions which have no outputs.
    This case happens if the original Rust function only takes *shared* borrows
    as inputs, and is thus pretty common.

    We are allowed to do this only because in this specific case,
    the backward function fails *exactly* when the forward function fails
    (they actually do exactly the same thing, the only difference being
    that the forward function can potentially return a value), and upon
    reaching the place where we should introduce a call to the backward
    function, we know we have introduced a call to the forward function.

    Also note that in general, backward functions "do more things" than
    forward functions, and have more opportunities to fail (even though
    in the generated code, calls to the backward functions should fail
    exactly when the corresponding, previous call to the forward functions
    failed).

    This optimization is done in {!SymbolicToPure}.  We might want to move it to
    the micro-passes subsequent to the translation from symbolic to pure, but it
    is really super easy to do it when going from symbolic to pure. Note that
    we later filter the useless *forward* calls in the micro-passes, where it is
    more natural to do.

    See the comments for {!PureMicroPasses.expression_contains_child_call_in_all_paths}
    for additional explanations.
 *)
let filter_useless_monadic_calls = ref true

(** If {!filter_useless_monadic_calls} is activated, some functions
    become useless: if this option is true, we don't extract them.

    The calls to functions which always get filtered are:
    - the forward functions with unit return value
    - the backward functions which don't output anything (backward
      functions coming from rust functions with no mutable borrows
      as input values - note that if a function doesn't take mutable
      borrows as inputs, it can't return mutable borrows; we actually
      dynamically check for that).
 *)
let filter_useless_functions = ref true

(** Simplify the forward/backward functions, in case we merge them
    (i.e., the forward functions return the backward functions).

    The simplification occurs as follows:
    - if a forward function returns the unit type and has non-trivial backward
      functions, then we remove the returned output.
    - if a backward function doesn't have inputs, we evaluate it inside the
      forward function and don't wrap it in a result.

    Example:
    {[
      // LLBC:
      fn incr(x: &mut u32) { *x += 1 }

      // Translation without simplification:
      let incr (x : u32) : result (unit * result u32) = ...
                                   ^^^^   ^^^^^^
                                    |     remove this result
                                remove the unit

      // Translation with simplification:
      let incr (x : u32) : result u32 = ...
    ]}
 *)
let simplify_merged_fwd_backs = ref true

(** Use short names for the record fields.

    Some backends can't disambiguate records when their field names have collisions.
    When this happens, we use long names, by which we concatenate the record
    names with the field names, and check whether there are name collisions.

    For backends which can disambiguate records (typically by using the typing
    information), we use short names (i.e., the original field names).
 *)
let record_fields_short_names = ref false

(** Parameterize the traits with their associated types, so as not to use
    types as first class objects.

    This is useful for some backends with limited expressiveness like HOL4,
    and to account for type constraints (like [fn f<T : Foo>(...) where T::bar = usize]).
 *)
let parameterize_trait_types = ref false

(** For sanity check: type check the generated pure code (activates checks in
    several places).

    TODO: deactivated for now because we need to implement the normalization of
    trait associated types in the pure code.
 *)
let type_check_pure_code = ref false

(** Shall we fail hard if we encounter an issue, or should we attempt to go
    as far as possible while leaving "holes" in the generated code? *)
let fail_hard = ref true

(** If true, add the type name as a prefix
    to the variant names.
    Ex.:
    In Rust:
      {[
        enum List = {
          Cons(u32, Box<List>),x
          Nil,
        }
      ]}

    F*, if option activated:
      {[
        type list =
        | ListCons : u32 -> list -> list
        | ListNil : list
      ]}

    F*, if option not activated:
      {[
        type list =
        | Cons : u32 -> list -> list
        | Nil : list
      ]}
 *)
let variant_concatenate_type_name = ref true

(** If true, extract the structures with unnamed fields as tuples.

    ex.:
    {[
      // Rust
      struct Foo(u32)

      // OCaml
      type Foo = (u32)
    ]}
 *)
let use_tuple_structs = ref true

let backend_has_tuple_projectors () =
  match !backend with Lean -> true | Coq | FStar | HOL4 -> false