Lossy Unification

The option --lossy-unification enables an experimental heuristic in the unification checker intended to improve its performance for unification problems of the form f es₀ = f es₁, i.e. unifying two applications of the same defined function, here f, to possibly different lists of arguments and projections es₀ and es₁. The heuristic is sound but not complete. In particular if Agda accepts code with the flag enabled it should also accept it without the flag (with enough resources, and possibly needing extra type annotations).

The option can be used either globally or in an OPTIONS pragma, in the latter case it applies only to the current module. There is also a pragma {-# INJECTIVE_FOR_INFERENCE f #-} which enables lossy unification only for f.

Heuristic

When trying to solve the unification problem f es₀ = f es₁ the heuristic proceeds by trying to solve es₀ = es₁, if that succeeds the original problem is also solved, otherwise unification proceeds as without the flag, likely by reducing both f es₀ and f es₁.

Example

Suppose f adds 100 to its input as defined below

f :   ℕ
f n = 100 + n

then to unify f 2 and f (1 + 1) the heuristic would proceed by unifying 2 with (1 + 1), which quickly succeeds. Without the flag we might instead first reduce both f 2 and f (1 + 1) to 102 and then compare those results.

The performance will improve most dramatically when reducing an application of f would produce a large term, perhaps an element of a record type with several fields and/or large embedded proof terms.

Drawbacks

One drawback is that in some cases performance of unification will be worse with the heuristic. Specifically, if the heuristic will repeatedly attempt to unify lists of arguments es₀ = es₁ while failing.

The main drawback is that the heursitic is not complete, i.e. it will cause Agda to ignore some possible solutions to unification variables. For example if f is a constant function, then the constraint f ?0 = f 1 does not uniquely determine ?0, but the heuristic will end up assigning 1 to ?0. However, if f is injective this heuristic is complete.

Such assignments can lead to Agda to report a type error which would not have been reported without the heuristic. This is because committing to ?0 = 1 might make other constraints unsatifiable.

Such assignments might also confuse readers. Note that with non-lossy unification you have the guarantee (in the absence of bugs) that, if the code type-checks, and you can find one consistent way to instantiate all meta-variables, then that is the way that the code is interpreted by the system. With lossy unification the solution you have in mind might not be the one the system uses.

Consider the following code, which is based on an example from López Juan’s PhD thesis (see Listing 6.16):

{-# OPTIONS --lossy-unification #-}
open import Agda.Builtin.Bool
open import Agda.Builtin.Nat

private variable
  m n : Nat

infixr 5 _∷_ _++_

data Bit-vector : Nat  Set where
  []  : Bit-vector 0
  _∷_ : Bool  Bit-vector n  Bit-vector (suc n)

_++_ : Bit-vector m  Bit-vector n  Bit-vector (m + n)
[]       ++ ys = ys
(x  xs) ++ ys = x  (xs ++ ys)

replicate :  n  Bool  Bit-vector n
replicate zero    x = []
replicate (suc n) x = x  replicate n x

vector : Bit-vector (m + n)
vector {m = m} {n = n} = replicate m true ++ replicate n false

Can you confidently predict the values of all of the following four bit-vectors? Are you sure that readers of your code can do this?

ex₁ : Bit-vector (0 + 1)
ex₁ = vector

ex₂ : Bit-vector (1 + 0)
ex₂ = vector

ex₃ : Bit-vector ((0 + 1) + (1 + 0))
ex₃ = xs ++ xs
  where
  xs = vector

ex₄ : Bit-vector ((0 + 1) + (1 + 0))
ex₄ = xs ++ xs
  where
  xs = vector {m = _}

References

Slow typechecking of single one-line definition, issue (#1625).