Version 7 (modified by diatchki, 2 years ago) (diff) |
---|

## Connection with GHC's Constraint Solver

The solver for the type nats is implemented as an
extra stage in GHC's constrraint solver (see `TcInteract.thePipeline`).

## Generating Evidence

The solver produces evidence (i.e., proofs) when computing new "given"
constraints, or when solving existing "wanted" constraints.
The evidence is constructed by applications of a set of pre-defined
rules. The rules are values of type `TypeRep.CoAxiomRule`.
Conceptually, rules have the form:

name :: forall tyvars. assumptions => conclusion

The rules have the usual logical meaning: the variables are universally quantified, and the assumptions imply the concluson. As a concrete example, consider the rule for left-cancellation of addtion:

AddCanceL :: forall a b c d. (a + b ~ d, a + c ~ d) => b ~ c

The type `CoAxiomRule` also supports infinte literal-indexed families
of simple axioms using constructor `CoAxiomTyLit`. These have the form:

name(l_1 .. l_n) :: conclusion

In this case `conclusion` is an equation that contains no type variables
but may depend on the literals in the name of the family. For example,
the basic definitional axiom for addition, `TcTypeNatsRules.axAddDef`,
uses this mechanism:

AddDef(2,3) :: 2 + 3 ~ 5

At present, the assumptions and conclusion of all rules are equations between types but this restriction is not important and could be lifted in the future.

The rules used by the solver are in module `TcTypeNatsRules`.

## The Solver

The entry point to the solver is `TcTypeNats.typeNatStage`.

We start by examining the constraint to see if it is obviously unsolvable
(using function `impossible`), and if so we stash it in the
constraint-solver's state and stop. Note that there is no assumption that
`impossible` is complete, but it is important that it is sound, so
if `impossible` returns `True`, then the constraint is definitely unsolvable,
but if `impossible` returns `False`, then we don't know if the constraint
is solvable or not.

The rest of the stage proceeds depending on the type of constraint, as follows.

### Given Constraints

Given constraints correspond to adding new assumptions that may be used
by the solver. We start by checking if the new constraint is trivial
(using function `solve`). A constraint is considered to be trivial
if it matches an already existing constraint or a rule that is known
to the solver. Such given constraints are ignored because they do not
contribute new information. If the new given is non-trivial, then it
will be recorded to the inert set as a new fact, and we proceed
to "interact" it with existing givens, in the hope of computing additional
useful facts (function `computeNewGivenWork`).

IMPORTANT: We assume that "given" constraints are processed before "wanted" ones. A new given constraint may be used to solve any existing wanted, so every time we added a new given to the inert set we should move all potentially solvable "wanted" constraint from the inert set back to the work queue. We DON'T do this, because it is quite inefficient: there is no obvious way to compute which "wanted"s might be affected, so we have to restart all of them!

The heart of the interaction is the function `interactCt`, which
performs one step of "forward" reasoning. The idea is to compute
new constraints whose proofs are made by an application of a rule
to the new given, and some existing givens. These new constraints are
added as new work, to be processed further on the next iteration of
GHC's constraint solver.

Aside: when we compute the new facts, we check to see if any are obvious contradictions. This is not strictly necessary because they would be detected on the next iteration of the solver. However, by doing the check early we get slightly better error messages because we can report the original constraint as being unsolvable (it leads to a contradiction), which tends to be easier to relate to the original program. Of course, this is not completely fool-proof---it is still possible that a contradiction is detected at a later iteration. An alternative idea---not yet implemented---would be to examine the proof of a contradiction and extract the original constraints that lead to it in the first place.

### Derived Constraints

`Derived`` constraints are facts that are implied by the constraints
in the inert set. They do not have complete proofs because
they may depend on proofs of as yet unsolved wanted constraints.
GHC does not associate any proof terms with derived constraints (to keep things simple?).
In the constraint solver, they are mostly used as "hints". For example,
consider the wanted constraint ``5 + 3 ~ x`, where `x` is a
free unification variable. These are the steps we'll take to solve
the constraint:

Rules: Add_def(5,3) : 5 + 3 ~ 8 Add_fun : forall a b c1 c2. (a + b ~ c1, a + b ~ c2) => c1 ~ c2 1. Add to inert set: [W] C: 5 + 3 ~ x 2. Generate new derived: [D] Add_fun(C,Add_def) : x ~ 8 (proof discarded) 3. GHC uses this hint to improve and reconsider the wanted: [W] C: 5 + 3 ~ 8 4. Solved: [W] C = Add_def(5,3)

The type-nat solver processes derived constraints in a similar fashion
to given constraints (`computeNewDerivedWork`): it checks to see if they are trivially known
and, if not, then it tries to generate some additional derived constraints.
The main difference is that derived constraints can be interacted
with all existing constraints to produce new facts, while given
constraints only interact with other givens.

### Wanted Constraints

The main purpose of the solver is to discharge `wanted`` constraints
(the purpose of processing given and derived constraints is to help
solve existing wanted goals). When we encounter a new wanted goals
we proceed as follows:
`

- Try to solve the goal, using a few different strategies:
- Try to see if it matches the conclusion of an iff rule (
`solveIff`). Aassumptions of rule become new wanted work. - Try to see if it matches an axiom exactly (
`solve`) - Try the ordering solver for
`<=`goals (`solveLeq`) - Try to use a (possibly synthesized) assumption

- Try to see if it matches the conclusion of an iff rule (

- If that didn't work:
- Wanted is added to the inert set
- Check to see if any of the existing wanteds in the inert set can be solved in terms of the new goal (
`reExamineWanteds`) - Generate new derived facts.