Elpi provides an easy-to-embed implementation of a dialect of λProlog, a programming language well suited to manipulate abstract syntax trees containing binders and unification variables.
Coq-Elpi provides a Coq plugin that lets one define new commands and tactics in
Elpi. For that purpose it provides an embedding of Coq's terms into λProlog
using the Higher-Order Abstract Syntax approach
(HOAS). It also
exports to Elpi a comprehensive set of Coq's primitives, so that one can
print a message, access the environment of theorems and data types, define a
new constant, declare implicit arguments, type classes instances, and so on.
For convenience it also provides quotations and anti-quotations for Coq's
syntax, so that one can write {{ nat -> lp:X }}
in the middle of a λProlog
program instead of the equivalent AST.
In the short term, provide an extension language for Coq well suited to manipulate terms containing binders. One can already use Elpi to implement commands and tactics. As ongoing research we are looking forward to express algorithms like higher order unification and type inference, and to provide an alternative elaborator for Coq.
The simplest way is to use OPAM and type
opam repo add coq-released https://coq.inria.fr/opam/released
opam install coq-elpi
The recommended user interface is VSCoq. We provide an extension for vscode in the market place, just look for Coq Elpi. The extension provides syntax hilighting for both languages even when they are nested via quotations and antiquotations.
Other editors (click to expand)
At the time of writing Proof General does not handle quotations correctly, see ProofGeneral/PG#437.
In particular Elpi Accumulate lp:{{ .... }}.
is used in tutorials to mix Coq and Elpi code
without escaping. Coq-Elpi also accepts Elpi Accumulate " .... ".
but strings part of the
Elpi code needs to be escaped. Finally, for non-tutorial material, one can always put
the code in an external file declared with From some.load.path Extra Dependency "filename" as f.
and use Elpi Accumulate File f.
.
CoqIDE does handle quotations. The installation process puts
coq-elpi.lang
in a place where CoqIDE can find it. Then you can select coq-elpi
from the menu Edit -> Preferences -> Colors
.
For Vim users, Coqtail provides syntax highlighting and handles quotations.
Development version (click to expand)
To install the development version one can type
opam pin add coq-elpi https://github.com/LPCIC/coq-elpi.git
One can also clone this repository and type make
, but check you have
all the dependencies installed first (see coq-elpi.opam).
We recommend to look at the CI setup for
ocaml versions being tested. Also, we recommend to install dot-merlin-reader
and ocaml-lsp-server
(version 1.15).
- The Elpi programming language is an Elpi tutorial, there is nothing Coq specific in there even if the tutorial uses Coq to step trough the various examples. If you never heard of λProlog or HOAS based languages (like Twelf or Beluga) then you are strongly encouraged to read this tutorial and have a look at λProlog's home page for additional documentation. Even if you are familiar with λProlog or HOAS it may be worth reading the last sections since they focus on Elpi specific features. Last but not least it covers common pitfalls for people with a background in functional programming and the tracing mechanisms (useful for debugging).
- HOAS of Coq terms focuses on how Coq terms are represented in Elpi, how to inspect them and call Coq APIs under a context of binders, and finally how holes ("evars" in Coq slang) are represented. It assumes the reader is familiar with Elpi.
- Writing commands in Elpi focuses on how to write commands, in particular how to store a state across calls via so called DBs and how to handled command arguments. It assumes the reader is familiar with Elpi and the HOAS of Coq terms.
- Writing tactics in Elpi describes how goals and tactics are represented, how to handle tactic arguments and finally how to define tactic notations. It assumes the reader is familiar with Elpi and the HOAS of Coq terms.
- Coq-Elpi in 20 minutes video recording of a talk given at the Coq Users and Developers Workshop 2020.
- reification is the typical use case for meta programs: reading the syntax of terms into an inductive representing a sub language on which some decision procedure can be implemented
- data bases shows how Elpi programs can store data and reuse it across multiple runs
- record expansion sketches a program to unpack records in a definition: it replaces an abstraction over a records with abstractions over all of its components
- record to sigma sketches a program that de-sugars a record type to iterated sigma types
- fuzzer sketches a program to alter an inductive type while preserving its well typedness. It makes nothing useful per se, but shows how to map a term and call the type checker deep inside it.
- tactics show how to create simple tactics by using (proof) terms and the elaborator of Coq
- generalize show how to abstract subterms out (one way to skin the cat, there are many)
- abs_evars show how to close a term containing holes (evars) with binders
- record import gives short names to record projections applied to the given record instance.
- reduction surgery implements a tactic fine tuning cbv with a list of allowed unfoldings taken from a module.
- Derive shows how to obtain proved equality tests and a few extra gadgets out of inductive type declarations. See the README for the list of derivations. It comes bundled with Coq-Elpi.
- Locker lets one hide the computational contents of definitions via modules or opaque locks. It comes bundled with Coq-Elpi.
- Hierarchy Builder is a Coq extension to declare hierarchies of algebraic structures.
- Algebra Tactics is a
port of the
ring
andfield
tactics to the Mathematical Components library. - Trakt is a generic goal preprocessing tool for proof automation tactics in Coq.
- Namespace Emulation System implements most of the features of namespaces (on top of Coq's modules).
- Dx uses elpi to generate an intermediate representation of Coq terms, to be later tranformed into C.
- Coercion enable to program coercions in Elpi. It comes bundled with Coq-Elpi.
In order to load Coq-Elpi use From elpi Require Import elpi
.
(click to expand)
Elpi Command <qname>
creates command named<qname>
containing the preamble elpi-command.Elpi Tactic <qname>
creates a tactic<qname>
containing the preamble elpi-tactic.Elpi Db <dbname> <code>
creates a Db (a program that is accumulated into other programs).<code>
is the initial contents of the Db, including the type declaration of its constituting predicates. It understands the#[phase]
attribute, see synterp-vs-interp.Elpi Program <qname> <code>
lower level primitive letting one crate a command/tactic with a custom preamble<code>
.From some.load.path Extra Dependency <filename> as <fname>
.Elpi Accumulate [<dbname>|<qname>] [<code>|File <fname>|Db <dbname>]
adds code to the current program (or<dbname>
or<qname>
if specified). The code can be verbatim, from a file or a Db. File names<fname>
must have been previously declared with the above command. It understands the#[skip="rex"]
and#[only="rex"]
which make the command a no op if the Coq version is matched (or not) by the given regular expression. It understands the#[phase]
attribute, see synterp-vs-interpElpi Typecheck [<qname>]
typechecks the current program (or<qname>
if specified). It understands the#[phase]
attribute, see synterp-vs-interpElpi Debug <string>
sets the variable<string>
, relevant for conditional clause compilation (the:if VARIABLE
clause attribute). It understands the#[phase]
attribute, see synterp-vs-interpElpi Trace [[<start> <stop>] <predicate-filter>*|Off]
enable/disable tracing, eventually limiting it to a specific range of execution steps or predicate names. It understands the#[phase]
attribute, see synterp-vs-interpElpi Trace Browser
enable/disable tracing for Elpi's trace browser.Elpi Bound Steps <number>
limits the number of steps an Elpi program can make.Elpi Print <qname> [<string> <filter>*]
prints the program<qname>
to an HTML file named<qname>.html
and a text file called<qname>.txt
(or<string>
if provided) filtering out clauses whose file or clause-name matches<filter>
. It understands the#[phase]
attribute, see synterp-vs-interp
where:
<qname>
is a qualified Coq name, e.g.derive.eq
ormy_program
.<dbname>
is like<qname>
but lives in a different namespace. By convention<dbname>
ends in.db
, e.g.derive.eq.db
.<code>
is verbatim Elpi code, eitherlp:{{ ... }}
or" ... "
(in the latter case, strings delimiters need to be escaped following Coq rules, e.g.lp:{{ coq.say "hello!" }}
becomes" coq.say ""hello!"" "
).<filename>
is a string containing the path of an external file, e.g."this_file.elpi"
.<start>
and<stop>
are numbers, e.g.17 24
.<predicate-filter>
is a regexp against which the predicate name is matched, e.g."derive.*"
.
(click to expand)
Since version 8.18 Coq has separate parsing and execution phases, respectively called synterp and interp.
Since Coq has an extensible grammar the parsing phase is not entirely
performed by the parser: after parsing one sentence Coq evaluates its
synterp action. The synterp actions of a command like Import A.
are
the subset of its effect which affect parsing, like enabling a notation.
Later, during the execution phase Coq evaluates the its
interp action, which includes effects like putting lemma names in scope or
enables type class instances etc.
Being able to parse an entire document quickly, without actually executing any sentence, is important for developing reactive user interfaces, but requires some extra work when defining new commands, in particular to separate their synterp actions from their interp ones. Each command defined with Coq-Elpi is split into two programs, one running during the parsing phase and the other one during the execution phase.
Each Elpi Command
internally declares two programs with the same name.
One to be run while the Coq document is parsed, the synterp-command,
and the other one while it is executed, the interp command.
Elpi Accumulate
, by default, adds code to the interp-command.
The #[phase]
attribute can be used to accumulate code to the synterp-command
or to both commands. Elpi Typecheck
checks both commands.
Each Elpi Db
internally declares one db, by default for the interp phase.
The #[phase]
attribute can be used crate a database for the synterp phase,
or for both phases. Note that databases for the two phases are distinct, no
data is shared among them. In particular the coq.elpi.accumulate*
API exists
in both phases and only acts on data bases for the current phase.
All synterp actions, i.e. calls to APIs dealing with modules and sections like begin/end-module or import/export, have to happen at both synterp and interp time and in the same order.
In order to do so, the synterp-command may need to communicate data to the corresponding interp-command. There are two ways for doing so.
The first one is to use, as the main entry points, the following ones:
pred main-synterp i:list argument, o:any.
pred main-interp i:list argument, i:any.
Unlike main
the former outputs a datum while the latter receives it in input.
During the synterp phase the API coq.synterp-actions
lists the actions
performed so far. An excerpt from the coq-builtin-synterp file:
% Action executed during the parsing phase (aka synterp)
kind synterp-action type.
type begin-module id -> synterp-action.
type end-module modpath -> synterp-action.
The synterp-command can output data of that type, but also any other data it wishes.
The second way to communicate data is implicit, but limited to synterp actions.
Such synterp actions can be recorded into (nested) groups whose structure is
declared using well-bracketed calls to predicates coq.begin-synterp-group
and coq.end-synterp-group
in the synterp phase. In the interp phase, one can
then use predicate coq.replay-synterp-action-group
to replay all the synterp
actions of the group with the given name at once.
In the case where one wishes to interleave code between the actions of a given
group, it is also possible to match the synterp group structure at interp, via
coq.begin-synterp-group
and coq.end-synterp-group
. Individual actions that
are contained in the group then need to be replayed individually.
One can use coq.replay-next-synterp-actions
to replay all synterp actions
until the next beginning/end of a synterp group. However, this is discouraged
in favour of using groups explicitly, as this is more modular. Code that used
to rely on the now-removed coq.replay-all-missing-synterp-actions
predicate
can rely on coq.replay-next-synterp-actions
instead, but this is discouraged
in favour of using groups explicitly)
#[phase="ph"]
where"ph"
can be"parsing"
,"execution"
or"both"
#[synterp]
is a shorthand for#[phase="parsing"]
#[interp]
is a shorthand for#[phase="execution]
(click to expand)
-
Elpi <qname> <argument>*.
invokes themain
predicate of the<qname>
program passing a possible empty list of arguments. This is how you invoke a command. -
elpi <qname> <argument>*.
invokes thesolve
predicate of the<qname>
program passing a possible empty list of arguments and the current goal. This is how you invoke a tactic. -
Elpi Export <qname> [As <other-qname>]
makes it possible to invoke command<qname>
(or<other-qname>
if given) without theElpi
prefix or invoke tactic<qname>
in the middle of a term just writing<qname> args
instead ofltac:(elpi <qname> args)
. Note that in the case of tactics, all arguments are considered to be terms. Moreover, remember that one can useTactic Notation
to give the tactic a better syntax and a shorter name when used in the middle of a proof script.
where <argument>
can be:
- a number, e.g.
3
, represented in Elpi as(int 3)
- a string, e.g.
"foo"
orbar.baz
, represented in Elpi as(str "foo")
and(str "bar.baz")
. Coq keywords and symbols are recognized as strings, eg=>
requires no quotes. Quotes are necessary if the string contains a space or a character that is not accepted for qualified identifiers or if the string isDefinition
,Axiom
,Record
,Structure
,Inductive
,CoInductive
,Variant
orContext
. - a term, e.g.
(3)
or(f x)
, represented in Elpi as(trm ...)
. Note that terms always require parentheses, that is3
is a number while(3)
is a Coq term and depending on the context could be a natural number (i.e.S (S (S O))
) or aZ
or ... See also the section Terms as arguments down below, and the syntax for Ltac variables down below.
Commands also accept the following arguments (the syntax is as close as possible
to the Coq one: [...] means optional, * means 0 or more). See the argument
data type in coq-builtin.elpi
for their HOAS encoding. See also the section
Terms as arguments down below.
Definition
name binder* [:
term]:=
termAxiom
name:
term- [
Record
|Structure
] name binder* [:
sort]:=
[name]{
name:
term;
*}
- [
Inductive
|CoInductive
|Variant
] name binder* [|
binder] [:
term]:=
|
name binder:
term * Context
binder*
Tactics also accept Ltac variables as follows:
ltac_string:(v)
(forv
of typestring
orident
)ltac_int:(v)
(forv
of typeint
orinteger
)ltac_term:(v)
(forv
of typeconstr
oropen_constr
oruconstr
orhyp
)ltac_(string|int|term)_list:(v)
(forv
of typelist
of ...)ltac_tactic:(t)
(fort
of typetactic_expr
)ltac_attributes:(v)
(forv
of typeattributes
) For example:
Tactic Notation "tac" string(X) ident(Y) int(Z) hyp(T) constr_list(L) simple_intropattern_list(P) :=
elpi tac ltac_string:(X) ltac_string:(Y) ltac_int:(Z) ltac_term:(T) ltac_term_list:(L) ltac_tactic:(intros P).
lets one write tac "a" b 3 H t1 t2 t3 [|m]
in any Ltac context.
Arguments are first interpreted by Ltac according to the types declared
in the tactic notation and then injected in the corresponding Elpi argument.
For example H
must be an existing hypothesis, since it is typed with
the hyp
Ltac type, but in Elpi it will appear as a term, eg trm c0
.
Similarly t1
, t2
and t3
are checked to be well typed and to contain no
unresolved implicit arguments, since this is what the constr
Ltac type means
If they were typed as open_constr
or uconstr
, the last or both checks would
be respectively skipped. In any case they are passed to the Elpi code as trm ...
.
Both "a"
and b
are passed to Elpi as str ...
.
Finally, ltac_term:(T)
and (T)
are not synonyms: but the former must be used
when defining tactic notations, the latter when invoking elpi tactics directly.
Attributes are supported in both commands and tactics. Examples:
#[ att ] Elpi cmd
#[ att ] cmd
for a commandcmd
exported viaElpi Export cmd
#[ att ] elpi tac
Tactic Notation ... attributes(A) ... := ltac_attributes:(A) elpi tac
. Due to a parsing conflict in Coq grammar, at the time of writing this code:has the following limitation:Tactic Notation "#[" attributes(A) "]" "tac" := ltac_attributes:(A) elpi tac.
#[ att ] tac.
does not parse(#[ att ] tac).
worksidtac; #[ att ] tac.
works
Since version 1.15, terms passed to Elpi commands code via (term)
or via a
declaration (like Record
, Inductive
...) are in elaborated format by
default. This means that all Coq notational facilities are available, like
deep pattern matching, or tactics in terms.
One can use the attribute #[arguments(raw)]
to declare a command which instead
takes arguments in raw format. In that case, notations are unfolded,
implicit arguments are expanded (holes _
are added) and lexical analysis is
performed (global names and bound names are identified, holes are applied
to bound names in scope), but deep pattern matching or tactics in terms are not
supported, and in particular type checking/inference is not performed.
Once can use the coq.typecheck
or coq.elaborate-skeleton
APIs
to fill in implicit arguments and insert coercions on raw terms.
Terms passed to Elpi tactics via tactic notations can be forced to be elaborated
beforehand by declaring the parameters to be of type constr
or open_constr
.
Arguments of type uconstr
are passed raw.
Elpi Query [<qname>] <code>
runs<code>
in the current program (or in<qname>
if specified).Elpi Query [<qname>] <synterp-code> <interp-code>
runs<synterp-code>
in the current (synterp) program (or in<qname>
if specified) and<interp-code>
in the current program (or<qname>
).elpi query [<qname>] <string> <argument>*
runs the<string>
predicate (that must have the same signature of the default predicatesolve
).
(click to expand)
- functional core (fun, forall, match, application, let-in, sorts)
- evars (unification variables)
- single Inductive and CoInductive types (including parameters, non-uniform parameters, indexes)
- mutual Inductive and CoInductive types
- fixpoints
- mutual fixpoints
- cofixpoints
- primitive records
- primitive projections
- primitive integers
- primitive floats
- primitive arrays
- universe polymorphism
- modules
- module types
- functor application
- functor definition
(click to expand)
Checked boxes are available, unchecked boxes are planned, missing items are not planned. This is a high level list, for the details see coq-builtin.
- i/o: messages, warnings, errors, Coq version
- logical environment: read, write, locate
- dependencies between objects
- type classes database: read, write
- take over resolution
- canonical structures database: read, write
- take over resolution
- coercions database: read, write
- sections: open, close
- scope management: import, export
- hints: mode, opaque, resolve, strategy
- arguments: implicit, name, scope, simpl
- abbreviations: read, write, locate
- typing and elaboration
- unification
- reduction:
lazy
,cbv
,vm
,native
- flags for
lazy
andcbv
- flags for
- ltac1: bridge to call ltac1 code, mono and multi-goal tactics
- option system: get, set, add
- pretty printer: boxes, printing width
- attributes: read
- coq-builtin documents the HOAS encoding of Coq terms and the API to access Coq
- coq-builtin-synterp documents APIs to interact with Coq at parsing time
- elpi-buitin documents Elpi's standard library, you may look here for list processing code
- coq-lib provides some utilities to manipulate Coq terms; it is an addendum to coq-builtin
- elpi-command-template provides the pre-loaded code for
Elpi Command
(execution phase) - elpi-command-template-synterp provides the pre-loaded code for
Elpi Command
(parsing phase) - elpi-tactic-template provides the pre-loaded code for
Elpi Tactic
The code of the Coq plugin is at the root of the repository in the src, elpi and theories directories.
The apps directory contains client applications written in Coq-Elpi.