drup / dowsing Goto Github PK

$ dowsindex search "string -> string -> string" | grep "^Stdlib"
Stdlib__String.cat : string -> string -> string
Stdlib__Pervasives.(^) : string -> string -> string
Stdlib__Filename.concat : string -> string -> string
Stdlib__StringLabels.cat : string -> string -> string
Stdlib__Filename.chop_suffix : string -> string -> string
Stdlib__Pervasives.max : 'a -> 'a -> 'a
Stdlib__Pervasives.min : 'a -> 'a -> 'a
Stdlib__Fun.const : 'a -> 'b -> 'a
Stdlib__Pervasives.fst : 'a * 'b -> 'a
Stdlib__Pervasives.snd : 'a * 'b -> 'b
Stdlib__Obj.magic : 'a -> 'b
Stdlib__Pervasives.failwith : string -> 'a
Stdlib__Pervasives.invalid_arg : string -> 'a

I would have expected to see every functions duplicated with their corresponding Stdlib.<Module> instead of just Stdlib__<Module> (see #13)

Given most libraries nowadays use dune and are packed, every functions will be technically available twice:

• once with the full flatten name
• once with the wrapped module

e.g.

$ dowsindex search "string -> string -> string" | grep "Js_of_ocaml_compiler"
Js_of_ocaml_compiler.Stdlib.String.cat : string -> string -> string
Js_of_ocaml_compiler__Stdlib.String.cat : string -> string -> string
Js_of_ocaml_compiler.Stdlib.Filename.concat : string -> string -> string
Js_of_ocaml_compiler__Stdlib.Filename.concat : string -> string -> string
Js_of_ocaml_compiler.Stdlib.Filename.chop_suffix : string -> string -> string
Js_of_ocaml_compiler__Stdlib.Filename.chop_suffix :

Realistically only the second one is ever used so it would be very handy to detect duplicates and only show the wrapped modules. (an optional argument can be available for users who want the unfiltered output)

Software engineering

• Add tests
• Implement a proper notion of printing environment
• Use proper logging
• Plug together the unification and the database again

Science

• Make sure the returned unifiers are minimal
• Add rigid variables

Indexation:

• Head
• Nb of arguments ("Equal" or "At least")
• Variances of constants
• Term indexing:
  To Read:
  • http://wwwlehre.dhbw-stuttgart.de/~sschulz/PAPERS/Schulz2013-FVI.pdf
  • Fingerprint Indexing for Paramodulation and Rewriting

Les types « 'a -> 'a list -> 'a » et « 'x -> 'x -> 'x » ne devraient pas être unifiables (en l'état, le programme prétend y arriver avec { 'a = 'a list ; 'x = 'a list }).

Par l'axiome (Ass-0) et curryfication, les types "a -> 'a -> b" et "a -> b" devraient être unifiables avec { 'a = unit } ; de même pour "a -> 'a -> 'b" et "a -> b". Ce n'est pas le cas en l'état.

In your README.md you say:
"Database format
Serialization
To speed up database deserialization, we could use a format that allows random access in an mmap datastructure."
Why don't you use Berkely DB or something like this (a persistent on disk hash table) instead of mem maping a data structure?

L'unification boucle toujours, en particulier sur un « fold » ("('a -> 'b -> 'b) -> 'a list -> 'b -> 'b") mais aussi sur "int list -> 'a -> unit".

There are cases where we have to fall down to proper unification. In this cases, it would be nice if it was fast.
Ideally, the stat function would have an option to give us a "hall of fame" of the slowest queries. There are cases where it's expected, but I think we should collect examples that are really out of line.

• 'X * int -> 'X
'a * 'b * 'c * ('a -> 'd) * ('b -> 'e) * ('c -> 'f) -> ('d * 'e * 'f)

Currently, the normal form is not stable by substitution. Tuples are represented as sorted arrays, this property is not preserved by substitution. There are 3 possible ways to fix it:

1. Change the normal form and remove the constraint on the array to be sorted. This solution has an impact on the equivalence test, which currently relies on this property.
2. Do the sorting after the substitution. Could have an impact on speed
3. Use another representation that preserves the equivalence modulo AC.

Also, one need to check that replacing a variable in the head position by an arrow is currently correctly handle.

L'unification boucle encore avec, par exemple, le type « 'a -> a -> 'a ».

To avoid crash when reading the database from the disk, a magic number corresponding to the version of Dowsing could be added at the start of the file.

Par l'isomorphisme (Ass-0), les types "unit * a" et "a" devraient être unifiables. La normalisation des deux types donne "(a)" (singleton) et "a", qui ne sont pas unifiés actuellement. On peut s'en sortir en traitant spécifiquement le cas d'un singleton dans [Type.make_tuple].

Currently, the handling of unifiers is a bit rough:

1. Unifier simplification simply "inlines" all the unnamed variables. It could do a bit more, like simplify alpha-conversions.
2. Unnamed variables should be given fresh names before printing.
3. The size function only counts the number of variables to be substituted, not the real size of the substitution.

To understand better the behaviour of the program, several parts of the program should be benchmarked.

• the complete unification procedure
• each part of the unification loop
  • the normalisation
  • the solver
  • the reinjection of the solution
• the Hullot tree