catala/compiler/scopelang/from_desugared.ml

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

855 lines
34 KiB
OCaml
Raw Normal View History

(* This file is part of the Catala compiler, a specification language for tax
and social benefits computation rules. Copyright (C) 2020 Inria, contributor:
Denis Merigoux <denis.merigoux@inria.fr>
Licensed under the Apache License, Version 2.0 (the "License"); you may not
use this file except in compliance with the License. You may obtain a copy of
the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
License for the specific language governing permissions and limitations under
the License. *)
2020-12-14 19:00:42 +03:00
(** Translation from {!module: Desugared.Ast} to {!module: Scopelang.Ast} *)
2022-11-21 12:46:17 +03:00
open Catala_utils
open Shared_ast
module D = Desugared.Ast
(** {1 Expression translation}*)
type target_scope_vars =
| WholeVar of ScopeVar.t
| States of (StateName.t * ScopeVar.t) list
type ctx = {
decl_ctx : decl_ctx;
scope_var_mapping : target_scope_vars ScopeVar.Map.t;
var_mapping : (Desugared.Ast.expr, untyped Ast.expr Var.t) Var.Map.t;
}
let tag_with_log_entry
(e : untyped Ast.expr boxed)
(l : log_entry)
2022-11-21 12:46:17 +03:00
(markings : Uid.MarkedString.info list) : untyped Ast.expr boxed =
2023-04-03 11:56:13 +03:00
if !Cli.trace_flag then
Expr.eapp
(Expr.eop (Log (l, markings)) [TAny, Expr.pos e] (Marked.get_mark e))
[e] (Marked.get_mark e)
else e
let rec translate_expr (ctx : ctx) (e : Desugared.Ast.expr) :
untyped Ast.expr boxed =
2022-07-11 12:32:23 +03:00
let m = Marked.get_mark e in
match Marked.unmark e with
2022-08-25 17:08:08 +03:00
| ELocation (SubScopeVar (s_name, ss_name, s_var)) ->
(* When referring to a subscope variable in an expression, we are referring
to the output, hence we take the last state. *)
let new_s_var =
match ScopeVar.Map.find (Marked.unmark s_var) ctx.scope_var_mapping with
| WholeVar new_s_var -> Marked.same_mark_as new_s_var s_var
| States states ->
Marked.same_mark_as (snd (List.hd (List.rev states))) s_var
in
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.elocation (SubScopeVar (s_name, ss_name, new_s_var)) m
2022-08-25 17:08:08 +03:00
| ELocation (DesugaredScopeVar (s_var, None)) ->
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.elocation
(ScopelangScopeVar
(match
ScopeVar.Map.find (Marked.unmark s_var) ctx.scope_var_mapping
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
with
| WholeVar new_s_var -> Marked.same_mark_as new_s_var s_var
| States _ -> failwith "should not happen"))
m
2022-08-25 17:08:08 +03:00
| ELocation (DesugaredScopeVar (s_var, Some state)) ->
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.elocation
(ScopelangScopeVar
(match
ScopeVar.Map.find (Marked.unmark s_var) ctx.scope_var_mapping
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
with
| WholeVar _ -> failwith "should not happen"
| States states -> Marked.same_mark_as (List.assoc state states) s_var))
m
| ELocation (ToplevelVar v) -> Expr.elocation (ToplevelVar v) m
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
| EVar v -> Expr.evar (Var.Map.find v ctx.var_mapping) m
| EStruct { name; fields } ->
Expr.estruct name (StructField.Map.map (translate_expr ctx) fields) m
| EDStructAccess { name_opt = None; _ } ->
(* Note: this could only happen if disambiguation was disabled. If we want
to support it, we should still allow this case when the field has only
one possible matching structure *)
Errors.raise_spanned_error (Expr.mark_pos m)
"Ambiguous structure field access"
| EDStructAccess { e; field; name_opt = Some name } ->
let e' = translate_expr ctx e in
let field =
try
StructName.Map.find name
(IdentName.Map.find field ctx.decl_ctx.ctx_struct_fields)
with Not_found ->
(* Should not happen after disambiguation *)
Errors.raise_spanned_error (Expr.mark_pos m)
2023-03-31 12:47:44 +03:00
"Field %a does not belong to structure %a"
(Cli.format_with_style [ANSITerminal.yellow])
("\"" ^ field ^ "\"")
(Cli.format_with_style [ANSITerminal.yellow])
(Format.asprintf "\"%a\"" StructName.format_t name)
in
Expr.estructaccess e' field name m
| ETuple es -> Expr.etuple (List.map (translate_expr ctx) es) m
| ETupleAccess { e; index; size } ->
Expr.etupleaccess (translate_expr ctx e) index size m
| EInj { e; cons; name } -> Expr.einj (translate_expr ctx e) cons name m
| EMatch { e; name; cases } ->
Expr.ematch (translate_expr ctx e) name
(EnumConstructor.Map.map (translate_expr ctx) cases)
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
m
| EScopeCall { scope; args } ->
Expr.escopecall scope
(ScopeVar.Map.fold
(fun v e args' ->
Make scopes directly callable Quite a few changes are included here, some of which have some extra implications visible in the language: - adds the `Scope of { -- input_v: value; ... }` construct in the language - handle it down the pipeline: * `ScopeCall` in the surface AST * `EScopeCall` in desugared and scopelang * expressions are now traversed to detect dependencies between scopes * transformed into a normal function call in dcalc - defining a scope now implicitely defines a structure with the same name, with the output variables of the scope defined as fields. This allows us to type the return value from a scope call and access its fields easily. * the implications are mostly in surface/name_resolution.ml code-wise * the `Scope_out` struct that was defined in scope_to_dcalc is no longer needed/used and the fields are no longer renamed (changes some outputs; the explicit suffix for variables with multiple states is ignored as well) * one benefit is that disambiguation works just like for structures when there are conflicts on field names * however, it's now a conflict if a scope and a structure have the same name (side-note: issues with conflicting enum / struct names or scope variables / subscope names were silent and are now properly reported) - you can consequently use scope names as types for variables as well. Writing literals is not allowed though, they can only be obtained by calling the scope. Remaining TODOs: - context variables are not handled properly at the moment - error handling on invalid calls - tests show a small error message regression; lots of examples will need tweaking to avoid scope/struct name or struct fields / output variable conflicts - add a `->` syntax to make struct field access distinct from scope output var access, enforced with typing. This is expected to reduce confusion of users and add a little typing precision. - document the new syntax & implications (tutorial, cheat-sheet) - a consequence of the changes is that subscope variables also can now be typed. A possible future evolution / simplification would be to rewrite subscopes as explicit scope calls early in the pipeline. That could also allow to manipulate them as expressions (bind them in let-ins, return them...)
2022-10-21 16:47:17 +03:00
let v' =
match ScopeVar.Map.find v ctx.scope_var_mapping with
Make scopes directly callable Quite a few changes are included here, some of which have some extra implications visible in the language: - adds the `Scope of { -- input_v: value; ... }` construct in the language - handle it down the pipeline: * `ScopeCall` in the surface AST * `EScopeCall` in desugared and scopelang * expressions are now traversed to detect dependencies between scopes * transformed into a normal function call in dcalc - defining a scope now implicitely defines a structure with the same name, with the output variables of the scope defined as fields. This allows us to type the return value from a scope call and access its fields easily. * the implications are mostly in surface/name_resolution.ml code-wise * the `Scope_out` struct that was defined in scope_to_dcalc is no longer needed/used and the fields are no longer renamed (changes some outputs; the explicit suffix for variables with multiple states is ignored as well) * one benefit is that disambiguation works just like for structures when there are conflicts on field names * however, it's now a conflict if a scope and a structure have the same name (side-note: issues with conflicting enum / struct names or scope variables / subscope names were silent and are now properly reported) - you can consequently use scope names as types for variables as well. Writing literals is not allowed though, they can only be obtained by calling the scope. Remaining TODOs: - context variables are not handled properly at the moment - error handling on invalid calls - tests show a small error message regression; lots of examples will need tweaking to avoid scope/struct name or struct fields / output variable conflicts - add a `->` syntax to make struct field access distinct from scope output var access, enforced with typing. This is expected to reduce confusion of users and add a little typing precision. - document the new syntax & implications (tutorial, cheat-sheet) - a consequence of the changes is that subscope variables also can now be typed. A possible future evolution / simplification would be to rewrite subscopes as explicit scope calls early in the pipeline. That could also allow to manipulate them as expressions (bind them in let-ins, return them...)
2022-10-21 16:47:17 +03:00
| WholeVar v' -> v'
| States ((_, v') :: _) ->
(* When there are multiple states, the input is always the first
one *)
v'
| States [] -> assert false
Make scopes directly callable Quite a few changes are included here, some of which have some extra implications visible in the language: - adds the `Scope of { -- input_v: value; ... }` construct in the language - handle it down the pipeline: * `ScopeCall` in the surface AST * `EScopeCall` in desugared and scopelang * expressions are now traversed to detect dependencies between scopes * transformed into a normal function call in dcalc - defining a scope now implicitely defines a structure with the same name, with the output variables of the scope defined as fields. This allows us to type the return value from a scope call and access its fields easily. * the implications are mostly in surface/name_resolution.ml code-wise * the `Scope_out` struct that was defined in scope_to_dcalc is no longer needed/used and the fields are no longer renamed (changes some outputs; the explicit suffix for variables with multiple states is ignored as well) * one benefit is that disambiguation works just like for structures when there are conflicts on field names * however, it's now a conflict if a scope and a structure have the same name (side-note: issues with conflicting enum / struct names or scope variables / subscope names were silent and are now properly reported) - you can consequently use scope names as types for variables as well. Writing literals is not allowed though, they can only be obtained by calling the scope. Remaining TODOs: - context variables are not handled properly at the moment - error handling on invalid calls - tests show a small error message regression; lots of examples will need tweaking to avoid scope/struct name or struct fields / output variable conflicts - add a `->` syntax to make struct field access distinct from scope output var access, enforced with typing. This is expected to reduce confusion of users and add a little typing precision. - document the new syntax & implications (tutorial, cheat-sheet) - a consequence of the changes is that subscope variables also can now be typed. A possible future evolution / simplification would be to rewrite subscopes as explicit scope calls early in the pipeline. That could also allow to manipulate them as expressions (bind them in let-ins, return them...)
2022-10-21 16:47:17 +03:00
in
ScopeVar.Map.add v' (translate_expr ctx e) args')
args ScopeVar.Map.empty)
Make scopes directly callable Quite a few changes are included here, some of which have some extra implications visible in the language: - adds the `Scope of { -- input_v: value; ... }` construct in the language - handle it down the pipeline: * `ScopeCall` in the surface AST * `EScopeCall` in desugared and scopelang * expressions are now traversed to detect dependencies between scopes * transformed into a normal function call in dcalc - defining a scope now implicitely defines a structure with the same name, with the output variables of the scope defined as fields. This allows us to type the return value from a scope call and access its fields easily. * the implications are mostly in surface/name_resolution.ml code-wise * the `Scope_out` struct that was defined in scope_to_dcalc is no longer needed/used and the fields are no longer renamed (changes some outputs; the explicit suffix for variables with multiple states is ignored as well) * one benefit is that disambiguation works just like for structures when there are conflicts on field names * however, it's now a conflict if a scope and a structure have the same name (side-note: issues with conflicting enum / struct names or scope variables / subscope names were silent and are now properly reported) - you can consequently use scope names as types for variables as well. Writing literals is not allowed though, they can only be obtained by calling the scope. Remaining TODOs: - context variables are not handled properly at the moment - error handling on invalid calls - tests show a small error message regression; lots of examples will need tweaking to avoid scope/struct name or struct fields / output variable conflicts - add a `->` syntax to make struct field access distinct from scope output var access, enforced with typing. This is expected to reduce confusion of users and add a little typing precision. - document the new syntax & implications (tutorial, cheat-sheet) - a consequence of the changes is that subscope variables also can now be typed. A possible future evolution / simplification would be to rewrite subscopes as explicit scope calls early in the pipeline. That could also allow to manipulate them as expressions (bind them in let-ins, return them...)
2022-10-21 16:47:17 +03:00
m
| ELit
((LBool _ | LInt _ | LRat _ | LMoney _ | LUnit | LDate _ | LDuration _) as
l) ->
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.elit l m
| EAbs { binder; tys } ->
let vars, body = Bindlib.unmbind binder in
let new_vars = Array.map (fun var -> Var.make (Bindlib.name_of var)) vars in
let ctx =
List.fold_left2
(fun ctx var new_var ->
2022-08-25 17:08:08 +03:00
{ ctx with var_mapping = Var.Map.add var new_var ctx.var_mapping })
ctx (Array.to_list vars) (Array.to_list new_vars)
in
Expr.eabs (Expr.bind new_vars (translate_expr ctx body)) tys m
| EApp { f = EOp { op; tys }, m1; args } ->
let args = List.map (translate_expr ctx) args in
Operator.kind_dispatch op
~monomorphic:(fun op -> Expr.eapp (Expr.eop op tys m1) args m)
~polymorphic:(fun op -> Expr.eapp (Expr.eop op tys m1) args m)
~overloaded:(fun op ->
match
Operator.resolve_overload ctx.decl_ctx
(Marked.mark (Expr.pos e) op)
tys
with
| op, `Straight -> Expr.eapp (Expr.eop op tys m1) args m
| op, `Reversed ->
Expr.eapp (Expr.eop op (List.rev tys) m1) (List.rev args) m)
| EOp _ -> assert false (* Only allowed within [EApp] *)
| EApp { f; args } ->
Expr.eapp (translate_expr ctx f) (List.map (translate_expr ctx) args) m
| EDefault { excepts; just; cons } ->
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.edefault
(List.map (translate_expr ctx) excepts)
(translate_expr ctx just) (translate_expr ctx cons) m
| EIfThenElse { cond; etrue; efalse } ->
Expr.eifthenelse (translate_expr ctx cond) (translate_expr ctx etrue)
(translate_expr ctx efalse)
m
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
| EArray args -> Expr.earray (List.map (translate_expr ctx) args) m
| EEmptyError -> Expr.eemptyerror m
| EErrorOnEmpty e1 -> Expr.eerroronempty (translate_expr ctx e1) m
2020-12-14 19:00:42 +03:00
(** {1 Rule tree construction} *)
(** Intermediate representation for the exception tree of rules for a particular
scope definition. *)
type rule_tree =
| Leaf of Desugared.Ast.rule list
(** Rules defining a base case piecewise. List is non-empty. *)
| Node of rule_tree list * Desugared.Ast.rule list
(** [Node (exceptions, base_case)] is a list of exceptions to a non-empty
list of rules defining a base case piecewise. *)
2020-12-14 19:00:42 +03:00
(** Transforms a flat list of rules into a tree, taking into account the
priorities declared between rules *)
let def_to_exception_graph
(def_info : Desugared.Ast.ScopeDef.t)
(def : Desugared.Ast.rule RuleName.Map.t) :
Desugared.Dependency.ExceptionsDependencies.t =
let exc_graph = Desugared.Dependency.build_exceptions_graph def def_info in
Desugared.Dependency.check_for_exception_cycle def exc_graph;
exc_graph
let rule_to_exception_graph (scope : Desugared.Ast.scope) = function
| Desugared.Dependency.Vertex.Var (var, state) -> (
let scope_def =
Desugared.Ast.ScopeDef.Map.find
(Desugared.Ast.ScopeDef.Var (var, state))
scope.scope_defs
in
let var_def = scope_def.D.scope_def_rules in
match Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input with
| OnlyInput when not (RuleName.Map.is_empty var_def) ->
(* If the variable is tagged as input, then it shall not be redefined. *)
Errors.raise_multispanned_error
((Some "Incriminated variable:", Marked.get_mark (ScopeVar.get_info var))
:: List.map
(fun (rule, _) ->
( Some "Incriminated variable definition:",
Marked.get_mark (RuleName.get_info rule) ))
(RuleName.Map.bindings var_def))
"It is impossible to give a definition to a scope variable tagged as \
input."
| OnlyInput -> Desugared.Ast.ScopeDef.Map.empty
(* we do not provide any definition for an input-only variable *)
| _ ->
Desugared.Ast.ScopeDef.Map.singleton
(Desugared.Ast.ScopeDef.Var (var, state))
(def_to_exception_graph
(Desugared.Ast.ScopeDef.Var (var, state))
var_def))
| Desugared.Dependency.Vertex.SubScope sub_scope_index ->
(* Before calling the sub_scope, we need to include all the re-definitions
of subscope parameters*)
let sub_scope_vars_redefs_candidates =
Desugared.Ast.ScopeDef.Map.filter
(fun def_key scope_def ->
match def_key with
| Desugared.Ast.ScopeDef.Var _ -> false
| Desugared.Ast.ScopeDef.SubScopeVar (sub_scope_index', _, _) ->
sub_scope_index = sub_scope_index'
(* We exclude subscope variables that have 0 re-definitions and are
not visible in the input of the subscope *)
&& not
((match
Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input
with
| Desugared.Ast.NoInput -> true
| _ -> false)
&& RuleName.Map.is_empty scope_def.scope_def_rules))
scope.scope_defs
in
let sub_scope_vars_redefs =
Desugared.Ast.ScopeDef.Map.mapi
(fun def_key scope_def ->
let def = scope_def.Desugared.Ast.scope_def_rules in
let is_cond = scope_def.scope_def_is_condition in
match def_key with
| Desugared.Ast.ScopeDef.Var _ -> assert false (* should not happen *)
| Desugared.Ast.ScopeDef.SubScopeVar (sscope, sub_scope_var, pos) ->
(* This definition redefines a variable of the correct subscope. But
we have to check that this redefinition is allowed with respect
to the io parameters of that subscope variable. *)
(match
Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input
with
| Desugared.Ast.NoInput ->
Errors.raise_multispanned_error
(( Some "Incriminated subscope:",
Marked.get_mark (SubScopeName.get_info sscope) )
:: ( Some "Incriminated variable:",
Marked.get_mark (ScopeVar.get_info sub_scope_var) )
:: List.map
(fun (rule, _) ->
( Some "Incriminated subscope variable definition:",
Marked.get_mark (RuleName.get_info rule) ))
(RuleName.Map.bindings def))
"It is impossible to give a definition to a subscope variable \
not tagged as input or context."
| OnlyInput when RuleName.Map.is_empty def && not is_cond ->
(* If the subscope variable is tagged as input, then it shall be
defined. *)
Errors.raise_multispanned_error
[
( Some "Incriminated subscope:",
Marked.get_mark (SubScopeName.get_info sscope) );
Some "Incriminated variable:", pos;
]
"This subscope variable is a mandatory input but no definition \
was provided."
| _ -> ());
let exc_graph = def_to_exception_graph def_key def in
let var_pos = Desugared.Ast.ScopeDef.get_position def_key in
exc_graph, sub_scope_var, var_pos)
sub_scope_vars_redefs_candidates
in
List.fold_left
(fun exc_graphs (new_exc_graph, subscope_var, var_pos) ->
Desugared.Ast.ScopeDef.Map.add
(Desugared.Ast.ScopeDef.SubScopeVar
(sub_scope_index, subscope_var, var_pos))
new_exc_graph exc_graphs)
Desugared.Ast.ScopeDef.Map.empty
(List.map snd (Desugared.Ast.ScopeDef.Map.bindings sub_scope_vars_redefs))
let scope_to_exception_graphs (scope : Desugared.Ast.scope) :
Desugared.Dependency.ExceptionsDependencies.t Desugared.Ast.ScopeDef.Map.t =
let scope_dependencies =
Desugared.Dependency.build_scope_dependencies scope
in
Desugared.Dependency.check_for_cycle scope scope_dependencies;
let scope_ordering =
Desugared.Dependency.correct_computation_ordering scope_dependencies
in
List.fold_left
(fun exceptions_graphs scope_def_key ->
let new_exceptions_graphs = rule_to_exception_graph scope scope_def_key in
Desugared.Ast.ScopeDef.Map.union
(fun _ _ _ -> assert false (* there should not be key conflicts *))
new_exceptions_graphs exceptions_graphs)
Desugared.Ast.ScopeDef.Map.empty scope_ordering
let build_exceptions_graph (pgrm : Desugared.Ast.program) :
Desugared.Dependency.ExceptionsDependencies.t Desugared.Ast.ScopeDef.Map.t =
ScopeName.Map.fold
(fun _ scope exceptions_graph ->
let new_exceptions_graphs = scope_to_exception_graphs scope in
Desugared.Ast.ScopeDef.Map.union
(fun _ _ _ -> assert false (* key conflicts should not happen*))
new_exceptions_graphs exceptions_graph)
pgrm.program_scopes Desugared.Ast.ScopeDef.Map.empty
(** Transforms a flat list of rules into a tree, taking into account the
priorities declared between rules *)
let def_map_to_tree
(def : Desugared.Ast.rule RuleName.Map.t)
(exc_graph : Desugared.Dependency.ExceptionsDependencies.t) : rule_tree list
=
(* we start by the base cases: they are the vertices which have no
successors *)
let base_cases =
Desugared.Dependency.ExceptionsDependencies.fold_vertex
(fun v base_cases ->
if
Desugared.Dependency.ExceptionsDependencies.out_degree exc_graph v = 0
then v :: base_cases
else base_cases)
exc_graph []
in
2023-04-07 17:35:09 +03:00
let rec build_tree (base_cases : Desugared.Dependency.ExceptionVertex.t) :
rule_tree =
let exceptions =
Desugared.Dependency.ExceptionsDependencies.pred exc_graph base_cases
in
let base_case_as_rule_list =
List.map
2023-04-07 17:35:09 +03:00
(fun (r, _) -> RuleName.Map.find r def)
(RuleName.Map.bindings base_cases.rules)
in
match exceptions with
| [] -> Leaf base_case_as_rule_list
| _ -> Node (List.map build_tree exceptions, base_case_as_rule_list)
in
List.map build_tree base_cases
2022-08-25 17:08:08 +03:00
(** From the {!type: rule_tree}, builds an {!constructor: Dcalc.EDefault}
2020-12-14 19:00:42 +03:00
expression in the scope language. The [~toplevel] parameter is used to know
when to place the toplevel binding in the case of functions. *)
let rec rule_tree_to_expr
~(toplevel : bool)
~(is_reentrant_var : bool)
(ctx : ctx)
(def_pos : Pos.t)
(params : Desugared.Ast.expr Var.t list option)
(tree : rule_tree) : untyped Ast.expr boxed =
let emark = Untyped { pos = def_pos } in
let exceptions, base_rules =
match tree with Leaf r -> [], r | Node (exceptions, r) -> exceptions, r
in
2023-02-20 19:21:44 +03:00
(* because each rule has its own variables parameters and we want to convert
the whole rule tree into a function, we need to perform some alpha-renaming
of all the expressions *)
let substitute_parameter
(e : Desugared.Ast.expr boxed)
(rule : Desugared.Ast.rule) : Desugared.Ast.expr boxed =
match params, rule.Desugared.Ast.rule_parameter with
| Some new_params, Some (old_params_with_types, _) ->
2023-02-20 19:21:44 +03:00
let old_params, _ = List.split old_params_with_types in
let old_params = Array.of_list (List.map Marked.unmark old_params) in
2023-02-20 19:21:44 +03:00
let new_params = Array.of_list new_params in
let binder = Bindlib.bind_mvar old_params (Marked.unmark e) in
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Marked.mark (Marked.get_mark e)
@@ Bindlib.box_apply2
2023-02-20 19:21:44 +03:00
(fun binder new_param -> Bindlib.msubst binder new_param)
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
binder
2023-02-20 19:21:44 +03:00
(new_params |> Array.map Bindlib.box_var |> Bindlib.box_array)
| None, None -> e
| _ -> assert false
(* should not happen *)
in
let ctx =
match params with
| None -> ctx
2023-02-20 19:21:44 +03:00
| Some new_params ->
ListLabels.fold_left new_params ~init:ctx ~f:(fun ctx new_param ->
match Var.Map.find_opt new_param ctx.var_mapping with
| None ->
let new_param_scope = Var.make (Bindlib.name_of new_param) in
{
ctx with
var_mapping =
Var.Map.add new_param new_param_scope ctx.var_mapping;
}
| Some _ ->
(* We only create a mapping if none exists because
[rule_tree_to_expr] is called recursively on the exceptions of
the tree and we don't want to create a new Scopelang variable for
the parameter at each tree level. *)
ctx)
in
let base_just_list =
List.map
(fun rule -> substitute_parameter rule.Desugared.Ast.rule_just rule)
base_rules
in
let base_cons_list =
List.map
(fun rule -> substitute_parameter rule.Desugared.Ast.rule_cons rule)
base_rules
in
let translate_and_unbox_list (list : Desugared.Ast.expr boxed list) :
untyped Ast.expr boxed list =
List.map
(fun e ->
2022-07-13 16:25:11 +03:00
(* There are two levels of boxing here, the outermost is introduced by
the [translate_expr] function for which all of the bindings should
have been closed by now, so we can safely unbox. *)
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
translate_expr ctx (Expr.unbox e))
list
in
let default_containing_base_cases =
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.make_default
(List.map2
(fun base_just base_cons ->
Expr.make_default []
(* Here we insert the logging command that records when a decision
is taken for the value of a variable. *)
(tag_with_log_entry base_just PosRecordIfTrueBool [])
base_cons emark)
(translate_and_unbox_list base_just_list)
(translate_and_unbox_list base_cons_list))
(Expr.elit (LBool false) emark)
(Expr.eemptyerror emark) emark
in
let exceptions =
2022-12-02 14:07:26 +03:00
List.map
(rule_tree_to_expr ~toplevel:false ~is_reentrant_var ctx def_pos params)
2022-12-02 14:07:26 +03:00
exceptions
in
2020-11-27 18:27:10 +03:00
let default =
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
Expr.make_default exceptions
(Expr.elit (LBool true) emark)
default_containing_base_cases emark
2020-11-27 18:27:10 +03:00
in
match params, (List.hd base_rules).Desugared.Ast.rule_parameter with
| None, None -> default
| Some new_params, Some (ls, _) ->
2023-02-20 19:21:44 +03:00
let _, tys = List.split ls in
2020-11-27 18:27:10 +03:00
if toplevel then
(* When we're creating a function from multiple defaults, we must check
2022-12-02 14:07:26 +03:00
that the result returned by the function is not empty, unless we're
dealing with a context variable which is reentrant (either in the
caller or callee). In this case the ErrorOnEmpty will be added later in
the scopelang->dcalc translation. *)
2022-12-02 14:07:26 +03:00
let default =
if is_reentrant_var then default else Expr.eerroronempty default emark
2022-12-02 14:07:26 +03:00
in
2023-02-20 19:21:44 +03:00
Expr.make_abs
2023-02-20 19:21:44 +03:00
(new_params
|> List.map (fun x -> Var.Map.find x ctx.var_mapping)
|> Array.of_list)
default tys def_pos
2020-11-27 18:27:10 +03:00
else default
2020-12-14 19:00:42 +03:00
| _ -> (* should not happen *) assert false
2020-12-14 19:00:42 +03:00
(** {1 AST translation} *)
2020-12-14 19:00:42 +03:00
(** Translates a definition inside a scope, the resulting expression should be
2022-08-25 17:08:08 +03:00
an {!constructor: Dcalc.EDefault} *)
let translate_def
~(is_cond : bool)
~(is_subscope_var : bool)
(ctx : ctx)
(def_info : Desugared.Ast.ScopeDef.t)
(def : Desugared.Ast.rule RuleName.Map.t)
(params : (Uid.MarkedString.info * typ) list Marked.pos option)
(typ : typ)
(io : Desugared.Ast.io)
(exc_graph : Desugared.Dependency.ExceptionsDependencies.t) :
untyped Ast.expr boxed =
(* Here, we have to transform this list of rules into a default tree. *)
let top_list = def_map_to_tree def exc_graph in
let is_input =
match Marked.unmark io.Desugared.Ast.io_input with
| OnlyInput -> true
| _ -> false
in
let is_reentrant =
match Marked.unmark io.Desugared.Ast.io_input with
| Reentrant -> true
| _ -> false
in
let top_value : Desugared.Ast.rule option =
if is_cond && ((not is_subscope_var) || (is_subscope_var && is_input)) then
(* We add the bottom [false] value for conditions, only for the scope
where the condition is declared. Except when the variable is an input,
where we want the [false] to be added at each caller parent scope. *)
Some
(Desugared.Ast.always_false_rule
(Desugared.Ast.ScopeDef.get_position def_info)
params)
else None
in
if
RuleName.Map.cardinal def = 0
&& is_subscope_var
(* Here we have a special case for the empty definitions. Indeed, we could
use the code for the regular case below that would create a convoluted
default always returning empty error, and this would be correct. But it
gets more complicated with functions. Indeed, if we create an empty
definition for a subscope argument whose type is a function, we get
something like [fun () -> (fun real_param -> < ... >)] that is passed as
an argument to the subscope. The sub-scope de-thunks but the de-thunking
does not return empty error, signalling there is not reentrant variable,
because functions are values! So the subscope does not see that there is
not reentrant variable and does not pick its internal definition instead.
See [test/test_scope/subscope_function_arg_not_defined.catala_en] for a
test case exercising that subtlety.
To avoid this complication we special case here and put an empty error
for all subscope variables that are not defined. It covers the subtlety
with functions described above but also conditions with the false default
value. *)
&& not (is_cond && is_input)
(* However, this special case suffers from an exception: when a condition is
defined as an OnlyInput to a subscope, since the [false] default value
will not be provided by the calee scope, it has to be placed in the
caller. *)
then
let m = Untyped { pos = Desugared.Ast.ScopeDef.get_position def_info } in
let empty_error = Expr.eemptyerror m in
match params with
| Some (ps, _) ->
let labels, tys = List.split ps in
Expr.make_abs
(Array.of_list
(List.map (fun lbl -> Var.make (Marked.unmark lbl)) labels))
empty_error tys (Expr.mark_pos m)
| _ -> empty_error
else
rule_tree_to_expr ~toplevel:true ~is_reentrant_var:is_reentrant ctx
(Desugared.Ast.ScopeDef.get_position def_info)
(Option.map
(fun (ps, _) ->
(List.map (fun (lbl, _) -> Var.make (Marked.unmark lbl))) ps)
params)
(match top_list, top_value with
| [], None ->
(* In this case, there are no rules to define the expression and no
default value so we put an empty rule. *)
Leaf [Desugared.Ast.empty_rule (Marked.get_mark typ) params]
| [], Some top_value ->
(* In this case, there are no rules to define the expression but a
default value so we put it. *)
Leaf [top_value]
| _, Some top_value ->
(* When there are rules + a default value, we put the rules as
exceptions to the default value *)
Node (top_list, [top_value])
| [top_tree], None -> top_tree
| _, None ->
Node (top_list, [Desugared.Ast.empty_rule (Marked.get_mark typ) params]))
let translate_rule
ctx
(scope : Desugared.Ast.scope)
(exc_graphs :
Desugared.Dependency.ExceptionsDependencies.t Desugared.Ast.ScopeDef.Map.t)
= function
| Desugared.Dependency.Vertex.Var (var, state) -> (
let scope_def =
Desugared.Ast.ScopeDef.Map.find
(Desugared.Ast.ScopeDef.Var (var, state))
scope.scope_defs
in
let var_def = scope_def.D.scope_def_rules in
let var_params = scope_def.D.scope_def_parameters in
let var_typ = scope_def.D.scope_def_typ in
let is_cond = scope_def.D.scope_def_is_condition in
match Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input with
| OnlyInput when not (RuleName.Map.is_empty var_def) ->
assert false (* error already raised *)
| OnlyInput -> []
(* we do not provide any definition for an input-only variable *)
| _ ->
let scope_def_key = Desugared.Ast.ScopeDef.Var (var, state) in
let expr_def =
translate_def ctx scope_def_key var_def var_params var_typ
scope_def.Desugared.Ast.scope_def_io
(Desugared.Ast.ScopeDef.Map.find scope_def_key exc_graphs)
~is_cond ~is_subscope_var:false
in
let scope_var =
match ScopeVar.Map.find var ctx.scope_var_mapping, state with
| WholeVar v, None -> v
| States states, Some state -> List.assoc state states
| _ -> failwith "should not happen"
in
[
Ast.Definition
( ( ScopelangScopeVar
(scope_var, Marked.get_mark (ScopeVar.get_info scope_var)),
Marked.get_mark (ScopeVar.get_info scope_var) ),
var_typ,
scope_def.Desugared.Ast.scope_def_io,
Expr.unbox expr_def );
])
| Desugared.Dependency.Vertex.SubScope sub_scope_index ->
(* Before calling the sub_scope, we need to include all the re-definitions
of subscope parameters*)
let sub_scope =
SubScopeName.Map.find sub_scope_index scope.scope_sub_scopes
in
let sub_scope_vars_redefs_candidates =
Desugared.Ast.ScopeDef.Map.filter
(fun def_key scope_def ->
match def_key with
| Desugared.Ast.ScopeDef.Var _ -> false
| Desugared.Ast.ScopeDef.SubScopeVar (sub_scope_index', _, _) ->
sub_scope_index = sub_scope_index'
(* We exclude subscope variables that have 0 re-definitions and are
not visible in the input of the subscope *)
&& not
((match
Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input
with
| Desugared.Ast.NoInput -> true
| _ -> false)
&& RuleName.Map.is_empty scope_def.scope_def_rules))
scope.scope_defs
in
let sub_scope_vars_redefs =
Desugared.Ast.ScopeDef.Map.mapi
(fun def_key scope_def ->
let def = scope_def.Desugared.Ast.scope_def_rules in
let def_typ = scope_def.scope_def_typ in
let is_cond = scope_def.scope_def_is_condition in
match def_key with
| Desugared.Ast.ScopeDef.Var _ -> assert false (* should not happen *)
| Desugared.Ast.ScopeDef.SubScopeVar (_, sub_scope_var, var_pos) ->
(* This definition redefines a variable of the correct subscope. But
we have to check that this redefinition is allowed with respect
to the io parameters of that subscope variable. *)
(match
Marked.unmark scope_def.Desugared.Ast.scope_def_io.io_input
with
| Desugared.Ast.NoInput -> assert false (* error already raised *)
| OnlyInput when RuleName.Map.is_empty def && not is_cond ->
assert false (* error already raised *)
| _ -> ());
(* Now that all is good, we can proceed with translating this
redefinition to a proper Scopelang term. *)
let expr_def =
translate_def ctx def_key def scope_def.D.scope_def_parameters
def_typ scope_def.Desugared.Ast.scope_def_io
(Desugared.Ast.ScopeDef.Map.find def_key exc_graphs)
~is_cond ~is_subscope_var:true
in
let subscop_real_name =
SubScopeName.Map.find sub_scope_index scope.scope_sub_scopes
in
Ast.Definition
( ( SubScopeVar
( subscop_real_name,
(sub_scope_index, var_pos),
match
ScopeVar.Map.find sub_scope_var ctx.scope_var_mapping
with
| WholeVar v -> v, var_pos
| States states ->
(* When defining a sub-scope variable, we always define
its first state in the sub-scope. *)
snd (List.hd states), var_pos ),
var_pos ),
def_typ,
scope_def.Desugared.Ast.scope_def_io,
Expr.unbox expr_def ))
sub_scope_vars_redefs_candidates
in
let sub_scope_vars_redefs =
List.map snd (Desugared.Ast.ScopeDef.Map.bindings sub_scope_vars_redefs)
in
sub_scope_vars_redefs
@ [
Ast.Call
( sub_scope,
sub_scope_index,
Untyped
{ pos = Marked.get_mark (SubScopeName.get_info sub_scope_index) }
);
]
2020-12-14 19:00:42 +03:00
(** Translates a scope *)
let translate_scope
(ctx : ctx)
(scope : Desugared.Ast.scope)
(exc_graphs :
Desugared.Dependency.ExceptionsDependencies.t Desugared.Ast.ScopeDef.Map.t)
: untyped Ast.scope_decl =
let scope_dependencies =
Desugared.Dependency.build_scope_dependencies scope
in
Desugared.Dependency.check_for_cycle scope scope_dependencies;
let scope_ordering =
Desugared.Dependency.correct_computation_ordering scope_dependencies
in
let scope_decl_rules =
List.fold_left
(fun scope_decl_rules scope_def_key ->
let new_rules = translate_rule ctx scope exc_graphs scope_def_key in
scope_decl_rules @ new_rules)
[] scope_ordering
in
(* Then, after having computed all the scopes variables, we add the
assertions. TODO: the assertions should be interleaved with the
definitions! *)
let scope_decl_rules =
scope_decl_rules
@ List.map
(fun e ->
Swap boxing and annotations in expressions This was the only reasonable solution I found to the issue raised [here](https://github.com/CatalaLang/catala/pull/334#discussion_r987175884). This was a pretty tedious rewrite, but it should now ensure we are doing things correctly. As a bonus, the "smart" expression constructors are now used everywhere to build expressions (so another refactoring like this one should be much easier) and this makes the code overall feel more straightforward (`Bindlib.box_apply` or `let+` no longer need to be visible!) --- Basically, we were using values of type `gexpr box = naked_gexpr marked box` throughout when (re-)building expressions. This was done 99% of the time by using `Bindlib.box_apply add_mark naked_e` right after building `naked_e`. In lots of places, we needed to recover the annotation of this expression later on, typically to build its parent term (to inherit the position, or build the type). Since it wasn't always possible to wrap these uses within `box_apply` (esp. as bindlib boxes aren't a monad), here and there we had to call `Bindlib.unbox`, just to recover the position or type. This had the very unpleasant effect of forcing the resolution of the whole box (including applying any stored closures) to reach the top-level annotation which isn't even dependant on specific variable bindings. Then, generally, throwing away the result. Therefore, the change proposed here transforms - `naked_gexpr marked Bindlib.box` into - `naked_gexpr Bindlib.box marked` (aliased to `boxed_gexpr` or `gexpr boxed` for convenience) This means only 1. not fitting the mark into the box right away when building, and 2. accessing the top-level mark directly without unboxing The functions for building terms from module `Shared_ast.Expr` could be changed easily. But then they needed to be consistently used throughout, without manually building terms through `Bindlib.apply_box` -- which covers most of the changes in this patch. `Expr.Box.inj` is provided to swap back to a box, before binding for example. Additionally, this gives a 40% speedup on `make -C examples pass_all_tests`, which hints at the amount of unnecessary work we were doing --'
2022-10-06 20:13:45 +03:00
let scope_e = translate_expr ctx (Expr.unbox e) in
Ast.Assertion (Expr.unbox scope_e))
scope.Desugared.Ast.scope_assertions
in
2020-11-27 13:37:21 +03:00
let scope_sig =
ScopeVar.Map.fold
(fun var (states : Desugared.Ast.var_or_states) acc ->
match states with
| WholeVar ->
let scope_def =
Desugared.Ast.ScopeDef.Map.find
(Desugared.Ast.ScopeDef.Var (var, None))
scope.scope_defs
in
let typ = scope_def.scope_def_typ in
ScopeVar.Map.add
(match ScopeVar.Map.find var ctx.scope_var_mapping with
| WholeVar v -> v
| States _ -> failwith "should not happen")
(typ, scope_def.scope_def_io)
acc
| States states ->
(* What happens in the case of variables with multiple states is
interesting. We need to create as many Var entries in the scope
signature as there are states. *)
List.fold_left
(fun acc (state : StateName.t) ->
let scope_def =
Desugared.Ast.ScopeDef.Map.find
(Desugared.Ast.ScopeDef.Var (var, Some state))
scope.scope_defs
in
ScopeVar.Map.add
(match ScopeVar.Map.find var ctx.scope_var_mapping with
| WholeVar _ -> failwith "should not happen"
| States states' -> List.assoc state states')
(scope_def.scope_def_typ, scope_def.scope_def_io)
acc)
acc states)
scope.scope_vars ScopeVar.Map.empty
2020-11-27 13:37:21 +03:00
in
let pos = Marked.get_mark (ScopeName.get_info scope.scope_uid) in
2020-11-27 13:37:21 +03:00
{
Ast.scope_decl_name = scope.scope_uid;
Ast.scope_decl_rules;
Ast.scope_sig;
Ast.scope_mark = Untyped { pos };
2023-01-20 20:18:53 +03:00
Ast.scope_options = scope.scope_options;
2020-11-27 13:37:21 +03:00
}
2020-12-14 19:00:42 +03:00
(** {1 API} *)
let translate_program
(pgrm : Desugared.Ast.program)
(exc_graphs :
Desugared.Dependency.ExceptionsDependencies.t Desugared.Ast.ScopeDef.Map.t)
: untyped Ast.program =
(* First we give mappings to all the locations between Desugared and This
involves creating a new Scopelang scope variable for every state of a
Desugared variable. *)
let ctx =
(* Todo: since we rename all scope vars at this point, it would be better to
have different types for Desugared.ScopeVar.t and Scopelang.ScopeVar.t *)
ScopeName.Map.fold
(fun _scope scope_decl ctx ->
ScopeVar.Map.fold
(fun scope_var (states : Desugared.Ast.var_or_states) ctx ->
let var_name, var_pos = ScopeVar.get_info scope_var in
let new_var =
match states with
| Desugared.Ast.WholeVar ->
WholeVar (ScopeVar.fresh (var_name, var_pos))
| States states ->
let var_prefix = var_name ^ "_" in
let state_var state =
ScopeVar.fresh
(Marked.map_under_mark (( ^ ) var_prefix)
(StateName.get_info state))
in
States (List.map (fun state -> state, state_var state) states)
in
{
ctx with
scope_var_mapping =
ScopeVar.Map.add scope_var new_var ctx.scope_var_mapping;
})
scope_decl.Desugared.Ast.scope_vars ctx)
pgrm.Desugared.Ast.program_scopes
{
scope_var_mapping = ScopeVar.Map.empty;
var_mapping = Var.Map.empty;
decl_ctx = pgrm.program_ctx;
}
in
let ctx_scopes =
ScopeName.Map.map
(fun out_str ->
let out_struct_fields =
ScopeVar.Map.fold
(fun var fld out_map ->
let var' =
match ScopeVar.Map.find var ctx.scope_var_mapping with
| WholeVar v -> v
| States l -> snd (List.hd (List.rev l))
in
ScopeVar.Map.add var' fld out_map)
out_str.out_struct_fields ScopeVar.Map.empty
in
{ out_str with out_struct_fields })
pgrm.Desugared.Ast.program_ctx.ctx_scopes
in
let new_program_scopes =
ScopeName.Map.fold
(fun scope_name scope new_program_scopes ->
let new_program_scope = translate_scope ctx scope exc_graphs in
ScopeName.Map.add scope_name new_program_scope new_program_scopes)
pgrm.program_scopes ScopeName.Map.empty
in
{
Ast.program_topdefs =
TopdefName.Map.map
(fun (e, ty) -> Expr.unbox (translate_expr ctx e), ty)
pgrm.program_topdefs;
Ast.program_scopes = new_program_scopes;
program_ctx = { pgrm.program_ctx with ctx_scopes };
}