ExtractionNameTricks

Pierre Letouzey edited this page Oct 27, 2017 · 18 revisions

When converting some code from Coq to another target language (e.g. OCaml, Haskell, ...), the extraction mechanism does its best to preserve the names occurring in the code. This eases any interaction with the extracted code (call to extracted functions, review of their types, etc). But this isn't always possible. For instance:

  • The syntactic conventions may differ, for instance OCaml and Haskell require a inductive constructor to be capitalized, while Coq has no such constraint.
  • A Coq name could clash with a reserved keyword of the target language (e.g. val in OCaml).
  • A name x may refer in Coq to different things thanks to module qualification (A.B.x versus C.x), while the target language may lack such feature, or provide only part of it. For instance there's no inner modules in Haskell. And even OCaml is slightly more restrictive than Coq when it comes to fully qualified names (see below).

As a consequence, the extraction might need to perform a renaming. And this turns out to be quite tedious.

Monolithic renaming

The historical approach was simply to (un)capitalize as needed, and in case of clash with a keyword or previous name, to insert or increase a numeral postfix at the end of the name, e.g. f to f0, or f1 if f0 is already used, etc. This approach is still used in the so-called "monolithic" extraction (Recursive Extraction or Extraction "file") when no inner modules or functors are involved. In this case, the extracted code is a simple sequence of definitions, whose names haven't already been used before. - Pros: the generated names are close to their Coq counterparts and almost of the same length. - Cons: we cannot easily predict whether some name f will become f, f0 or f27, since this depends on the "context" of f after extraction, e.g. the amount of other f constants in other libraries that will be incorporated in the extracted code. In particular, this method cannot be used for extracting modules with signatures : there would be no guarantee that a name is translated similarly in the implementation and in the signature.

Modular renaming

Another renaming mode tries to produce predictable names. This mode is used for inner modules or functors, and when using the Extraction Library command (predictable names help when the extracted code is meant to be some library used by some other code). When creating an uncapitalized name, we insert the prefix coq_ before the original name when this name is capitalized, or is a keyword, or already starts with a coq_ prefix. The same is done for obtaining capitalized name, with the prefix Coq_. This renaming schema cannot lead to name clashes since it is bijective : from an obtained uncapitalized name we could retrieve the initial one by removing the 4 first characters if they are coq_ or doing nothing otherwise. Note that the prefixes we use aren't reserved : the user can start her Coq names with them, it's just a bad idea since extracted names will then be ugly. Also note that the prefixes we use aren't magical : we just tried to pick ones that would be rare in Coq developments while not being too long nor cryptic.

Haskell and qualified names

When extracting to Haskell, only the modules-as-v-files structure is kept. Applied functors are expanded (that's defunctorization). Inner modules are "flattened" : their content is placed at the toplevel of the current file. To avoid name clashes during these flattenings, the following renaming takes place : A.B.c become A__B__c, and we add an extra _ prefix if the generated name is capitalized while it should not be. Note that unlike the previous renamings, this isn't clash-safe : the user could have used a Coq name A__B__c.

  • For the moment, a warning informs the user that the substring __ is reserved for the extraction and shouldn't be use inside user names.
  • I don't see (yet?) a solution that would be a) predictable b) clash-free c) without reserved substring. For instance, adding a specific prefix in front of A__B__c might help, but that doesn't fully avoid names clashes: A.B.c and A__B.c would still be extracted to the same names. A possible compromise could be to transform the module path into a unique number (hash?), leading to names of the form coq_123_c. These names aren't predictable easily, but we might add comments relating these unique numbers and the corresponding module path ?

Other painful situations

For now, most of the solutions implemented to the issues below implies the use of the __ substring in a non-clash-safe way, and hence a warning informs the user that this substring is reserved for the extraction.

Co-induction (OCaml)

For a co-inductive type foo, we currently produce a mutual block in OCaml :

type foo = __foo Lazy.t and __foo = FooConstr1 ... | FooConstr2 ...

In the rest of the code, it's foo that appears : __foo is only an internal intermediate step. But it may well interfere with a user-defined type of the same name.

  • Alternative solution that would be clash-safe : use another prefix, for instance here use the name coqL__foo (with L as in Lazy).

Anonymous fields in records (OCaml)

In Coq, it is not mandatory to name all record fields.

Record foo := { field : nat; _ : bool}.

Currently, the extraction generates a field name of the form foo__123 for the 123-th field of record foo when this field is anonymous.

  • Alternative solution that would be clash-safe : use another prefix, for instance here use the name coqF__foo__123 (with F as in field).
  • Note that Haskell isn't concerned for the moment, since records are treated as standard inductive types during this extraction.

Modules and inaccessible qualified names (OCaml)

Even if the Coq and OCaml module systems are very similar, the first one is slightly more permissive when it comes to qualifying names.

Module M.
 Definition t := 0.
 Module N.
  Definition t := 1.
  Definition u := M.t.
 End N.
End M.

In this example, the module M can be used to refer to objects even though M isn't finished yet. OCaml doesn't allow that for normal modules. And recursive modules have significant restrictions : an interface should be present after a module rec M, and the translation to OCaml of the previous example (with a module rec M) is rejected (Error : cannot safely evaluate the definition of the recursively-defined module M). We currently handle this situation via extra local modules, whose generated names are of the form Coq__123. Here:

module M = struct
 module Coq__0 = struct
  let t = O
 end
 let t = Coq__0.t
 module N = struct
  let t = S O
  let u = Coq__0.t
 end
end
  • Note : this is actually clash-safe : the part after the Coq__ prefix is a number, and cannot interfere with identifiers coming from a modular rename (see section 2 above). And a user module of name Coq__22 would have been converted to Coq__Coq__22.
  • Note : these local disambiguating modules are also used in the following situation, which is rather common:
Module M.
 Definition t := nat.
 Module N.
  Definition t := t.
 End N.
End M.

In OCaml, a type t = t would be wrong, since type definition is recursive by default, so we also use here a local naming submodule. TODO: for OCaml >= 4.02, we could generate here type nonrec t = t.

Module parameters (OCaml)

Inside a functor, its parameter name might be shadowed, but the content of the parameter might remain accessible.

Module Type T. Parameter t : nat. End T.
Module F (X:T).
 Module X. Definition u := 0. End X.
 Definition v := X.t. (* This X is the parameter of F, not the inner submodule !! *)
End F.

To support this "peculiar" feature, the extraction renames here the functor parameter, using an internal unique id (see MBId in names.ml), leading to names such as X__123.

  • Alternative solution that would be type-safe : use another prefix, for instance here use the name coqP__X__123 (with P as in parameter).

Unicode characters

OCaml doesn't accept unicode characters in names while Coq does.

Definition αβγ := 123.

We currently convert each non-ascii characters to a substring of the form __U1234_ where 1234 is the unicode index of the character.

  • I see no proper solution that would be a) predictable b) clash-safe c) without reserved substring such as __. In particular, how to generate different names for and a__U03B1_ (with 03B1 being the unicode index of α) ? At least, we could consider here that the reserved substring is rather __U, which is probably much less frequent than __.
  • In Haskell, the most recent norm accept unicode letters in names, and apparently ghc supports it (but hugs does not). Anyway, before leaving unicode characters in Haskell extraction (probably via a user flag), we should first compare the range of unicode letters accepted by Coq and by Haskell.

First-level inner modules in monolithic mode (OCaml)

In the so-called "monolithic" mode, the content of many Coq files is assembled in one OCaml file. If these Coq files contain inner modules of the same name, we could end with a name clash. For the moment these 1st-level inner modules are renamed if necessary by prefixing them with Coq_, Coq0_, Coq1_, etc.

A final note

__ is also used by the extraction to name the constant which is substituted to eliminated code. With respect to renaming, this isn't a big deal: We simply consider this __ name as a keyword of the target language, so any user-defined __ is renamed into something else.

See also: Pull Request 149.

Clone this wiki locally
You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.
Press h to open a hovercard with more details.