| Age | Commit message (Collapse) | Author |
|---|
|
This assertion checked that two arguments in the same position were equal,
but in fact, since one might have already been reduced, they are only
convertible (which is too costly to check in an assertion).
|
|
|
|
The no-inversion and maximal abstraction over dependencies now
supports abstraction over goal variables rather than only on "rel"
variables. In particular, it now works consistently using
"intro H; refine (match H with ... end)" or
"refine (fun H => match H with ... end)".
By doing so, we ensure that all three strategies are tried in all
situations where a return clause has to be inferred, even in the
context of a "refine".
See antepenultimate commit for discussion.
|
|
even when no type constraint is given.
This no-inversion and maximal abstraction over dependencies in (rel)
variables heuristic was used only when a type constraint was given.
By doing so, we ensure that all three strategies "inversion with
dependencies as evars", "no-inversion and maximal abstraction over
dependencies in (rel) variables", "no-inversion and no abstraction
over dependencies" are tried in all situations where a return clause
has to be inferred.
See penultimate commit for discussion.
|
|
The no-inversion no-dependency heuristic was used only in the absence
of type constraint. We may now use it also in the presence of a type
constraint.
See previous commit for discussion.
|
|
As noted by Jason Gross on coq-club (Aug 18, 2016), the "small
inversion" heuristic is not used consistently depending on whether the
variables in the type constraint are Rel or Var.
This commit simply gives uniformly preference to the inversion of the
predicate along the indices of the type over other heuristics.
The next three commits will improve further a uniform use of the
different heuristics.
----------------------------------------------------------------------
Here are some extra comments on how to go further with the inference
of the return predicate:
The "small inversion" heuristic build_inversion_problem (1) is
characterized by two features:
- small inversion properly speaking (a), i.e. that is for a match on
t:I params p1(u11..u1p1) ... pn(un1..unpn) with pi exposing the
constructor structure of the indices of the type of t, a return
clause of the form "fun x1..xn (y:I params x1..xn) => match x1..xn y with
| p1(z11..z1p1) ... pn(zn1..znpn) => ?T@{z11..znpn}
| _ => IDProp
end" is used,
- the dependent subterms in the external type constraint U are replaced
by existential variables (b) which can be filled either by projecting
(i.e. installing a dependency) or imitating (i.e. no dependency);
this is obtained by solving the constraint ?T@{u11..unpn} == U by
setting ?T@{z11..znpn} := U'(...?wij@{zij:=uij}...) where U has been
written under the form U'(...uij...) highlighting all occurrences of
each of the uij occurring in U; otherwise said the problem is reduced to
the question of instantiating each wij, deciding whether wij@{zij} := zij
(projection) or wij@{zij} := uij (imitation) [There may be different
way to expose the uij in U, e.g. in the presence of overlapping, or of
evars in U; this is left undetermined].
The two other heuristics used are:
- prepare_predicate_from_arsign_tycon (2): takes the external type
constraint U and decides that each subterm of the form xi or y for a
match on "y:I params x1 ... xn" is dependent; otherwise said, it
corresponds to the degenerated form of (1) where
- no constructor structure is exposed (i.e. each pi is trivial)
- only uij that are Rel are replaced by an evar ?wij and this evar is
directly instantiated by projection (hence creating a dependency),
- simple use of of an evar in case no type constraint is given (3):
this evar is not dependent on the indices nor on the term to match.
Heuristic (1) is not strictly more powerful than other heuristics
because of (at least) two weaknesses.
- The first weakness is due to feature (b), i.e. to letting
unification decide whether these evars have to create a dependency
(projection) or not (imitation).
In particular, the heuristic (2) gives priority to systematic
abstraction over the dependencies (i.e. giving priority to
projection over imitation) and it can then be better as the
following example (from RelationClasses.v) shows:
Fixpoint arrows (l : Tlist) (r : Type) : Type :=
match l with
| Tnil => r
| A :: l' => A -> arrows l' r
end.
Fixpoint predicate_all (l : Tlist) : arrows l Prop -> Prop :=
match l with
| Tnil => fun f => f
| A :: tl => fun f => forall x : A, predicate_all tl (f x)
end.
Using (1) fails. It proposes the predicate
"fun l' => arrows ?l[l':=l'] Prop" so that typing the first branch
leads to unify "arrows ?l[l:=Tnil] Prop == Prop", a problem about
which evarconv unification is not able (yet!) to see what are the
two possible solutions. Using (2) works. It instead directly
suggests that the predicate is "fun l => arrows l Prop" is used, so
that unification is not needed.
Even if in practice the (2) is good (and hence could be added to
(1)), it is not universally better. Consider e.g.
y:bool,H1:P y,H2:P y,f:forall y, P y -> Q y |-
match y as z return Q y with
| true => f y H1
| false => f y H2
end : Q y
There is no way to type it with clause "as z return Q z" even if
trying to generalize H1 and H2 so that they get type P z.
- A second weakness is due to the interaction between small inversion
and constructors having a type whose indices havex a less refined
constructor structure than in the term to match, as in:
Inductive I : nat -> Set :=
| C1 : forall n : nat, listn n -> I n
| C2 : forall n : nat, I n -> I n.
Check (fun x : I 0 => match x with
| C1 n l => 0
| C2 n c => 0
end).
where the inverted predicate is "in I n return match n with 0 => ?T | _ => IDProp end"
but neither C1 nor C2 have fine enough types so that n becomes
constructed. There is a generic solution to that kind of situation which
is to compile the above into
Check (fun x : I 0 => match x with
| C1 n l => match n with 0 => 0 | _ -> id end
| C2 n c => match n with 0 => 0 | _ -> id end
end).
but this is not implemented yet.
In the absence of this refinement, heuristic (3) can here work
better.
So, the current status of the claim is that for (1) to be strictly
more powerful than other current heuristics, work has to be done
- (A) at the unification level (by either being able to reduce problems of
the form "match ?x[constructor] with ... end = a-rigid-term", or, at
worst, by being able to use the heuristic favoring projecting for such
a problem), so that it is better than (2),
- (B) at the match compilation level, by enforcing that, in each branch,
the corresponding constructor is refined so has to match (or
discriminate) the constraints given by the type of the term to
match, and hence being better than (3).
Moreover, (2) and (3) are disjoint. Here is an example which (3) can
solve but not (2) (and (1) cannot because of (B)). [To be fixed in
next commit.]
Inductive I : bool -> bool -> Type := C : I true true | D x : I x x.
Check fun z P Q (y:I true z) (H1 H2:P y) (f:forall y, P y -> Q y z) =>
match y with
| C => f y H1
| D _ => f y H2
end : Q y z.
Indeed, (2) infers "as y' in I b z return Q y z" which does not work.
Here is an example which (2) can solve but not (3) (and (1) cannot
because of (B) again). [To be fixed in 2nd next commit].
Check fun z P Q (y:I true z) (H1 H2:P y) (f:forall y z, P y -> Q y z) =>
match y with
| C => f y true H1
| D b => f y b H2
end : Q y z.
|
|
The fix is essentially a revert of f22ad60 that introduced the use of the
pretyper version of whd_all instead of the one from the kernel. The exact
cause of slowness of the pretyper version is still to be investigated, but
it is believed that it is due to a call-by-name strategy, to compare with
call-by-need in the kernel.
Note that there is still something fishy in presence of evars. Technically
vm_compute does not handle them, but if this comes to be the case, it might
be worthwile to use an evar-aware variant of whd_all.
|
|
|
|
|
|
|
|
We untangle many dependencies on Ltac datastructures and modules from the
lower strata, resulting in a self-contained ltac/ folder. While not a plugin
yet, the change is now very easy to perform. The main API changes have been
documented in the dev/doc/changes file.
The patches are quite rough, and it may be the case that some parts of the
code can migrate back from ltac/ to a core folder. This should be decided on
a case-by-case basis, according to a more long-term consideration of what is
exactly Ltac-dependent and whatnot.
|
|
|
|
|
|
|
|
|
|
Typing.type_of was using conversion for types of fixpoints while it
could have used unification.
|
|
Add a boolean for refolding during reduction, and an option
that is off by default in 8.6, to turn refolding on in all reduction
functions, as in 8.5.
|
|
This legacy function is still used by destruct, and is a hotspot in various
examples from the wild. We hijack the check from Typeclass and perform a
double check at once not to mark unresolvable evars in vain a lot.
|
|
This is a follow-up to #244 and corrects a minor bug introduced by the
patch.
|
|
|
|
|
|
|
|
|
|
composition operator.
Short story:
This pull-request:
(1) removes the definition of the "right-to-left" function composition operator
(2) adds the definition of the "left-to-right" function composition operator
(3) rewrites the code relying on "right-to-left" function composition to rely on "left-to-right" function composition operator instead.
Long story:
In mathematics, function composition is traditionally denoted with ∘ operator.
Ocaml standard library does not provide analogous operator under any name.
Batteries Included provides provides two alternatives:
_ % _
and
_ %> _
The first operator one corresponds to the classical ∘ operator routinely used in mathematics.
I.e.:
(f4 % f3 % f2 % f1) x ≜ (f4 ∘ f3 ∘ f2 ∘ f1) x
We can call it "right-to-left" composition because:
- the function we write as first (f4) will be called as last
- and the function write as last (f1) will be called as first.
The meaning of the second operator is this:
(f1 %> f2 %> f3 %> f4) x ≜ (f4 ∘ f3 ∘ f2 ∘ f1) x
We can call it "left-to-right" composition because:
- the function we write as first (f1) will be called first
- and the function we write as last (f4) will be called last
That is, the functions are written in the same order in which we write and read them.
I think that it makes sense to prefer the "left-to-right" variant because
it enables us to write functions in the same order in which they will be actually called
and it thus better fits our culture
(we read/write from left to right).
|
|
|
|
composition in a more obvious way
This commit rewrites terms
(fun x -> f1 (f2 ... (fN x)...))
to
f1 % f2 % ... % fN
|
|
"Context.Named.{to,of}_rel_decl"
|
|
|
|
mainly concerning referring to "Context.{Rel,Named}.get_{id,value,type}" functions.
If multiple modules define a function with a same name, e.g.:
Context.{Rel,Named}.get_type
those calls were prefixed with a corresponding prefix
to make sure that it is obvious which function is being called.
|
|
|
|
|
|
more cleanups
|
|
|
|
|
|
|
|
Instead of recomputing the evar name environment from scratch when it is
unchanged, we simply return the original one.
This should fix #4964 for good, although I still find the global evar naming
mechanism from Pretyping more than messy. Introducing the heuristic allowing
to capture variables from Ltac in constr binders is indeed the root of many
evils... That is far from being a zero-cost abstraction!
|
|
|
|
Suggested by @ppedrot
|
|
As noted by @ppedrot, the first is redundant. The patch is basically a renaming.
We didn't make the component optional yet, but this could happen in a
future patch.
|
|
In some cases prior to this patch, there were two cases for the same
error function, one taking a location, the other not.
We unify them by using an option parameter, in the line with recent
changes in warnings and feedback.
This implies a bit of clean up in some places, but more importantly, is
the preparation for subsequent patches making `Loc.location` opaque,
change that could be use to improve modularity and allow a more
functional implementation strategy --- for example --- of the
beautifier.
|
|
This is a followup to 91ee24b4a7843793a84950379277d92992ba1651 , where
we got a few cases wrong wrt to newline endings.
Thanks to @herbelin for pointing it out.
This doesn't yet fix https://coq.inria.fr/bugs/show_bug.cgi?id=4842
|
|
Delimit the scope of the failure to ease potential need for debugging
the debugging printer.
Protect against one of the causes of failure (calling
get_family_sort_of with non-synchronized sigma).
|
|
We used to recompute all fresh named contexts for evars before this patch in
the push_rel_context_to_named_context function. This was incurring a linear
penalty and a memory explosion due to the reallocation of many arrays. Now, we
rather remember the context between evar creations by sharing it in the pretyping
environment.
This can be considered as a fix for bug #4964 even though we might do better.
|
|
|
|
This saves a quadratic allocation by replacing arrays with maps.
|
|
In addition to sharing, we also delay the computation of the environment in
a by-need fashion.
|
|
|
|
A couple of bugs have been found.
Example #4932 is now printing correctly in the presence of multiple
binders (when no let-in, no irrefutable patterns).
|
|
|
|
|
