Agda-2.5.2: A dependently typed functional programming language and proof assistant

Safe HaskellNone
LanguageHaskell2010

Agda.TypeChecking.Reduce.Fast

Description

This module contains an optimised implementation of the reduction algorithm from Reduce and Match. It runs roughly an order of magnitude faster than the original implementation.

The differences are the following:

  • Only applies when we don't have --sharing and when all reductions are allowed.

This means we can skip a number of checks that would otherwise be performed at each reduction step.

  • Does not track whether simplifications were made.

This information is only used when trying to simplify terms, so the simplifier runs the slow implementation.

  • Precomputes primZero and primSuc.

Since all the functions involved in reduction are implemented in this module in a single where block, we can look up zero and suc once instead of once for each reduction step.

  • Run outside ReduceM

ReduceM is already just a plain reader monad, but pulling out the environment and doing all reduction non-monadically saves a significant amount of time.

  • Memoise getConstInfo.

A big chunk of the time during reduction is spent looking up definitions in the signature. Any long-running reduction will use only a handful definitions though, so memoising getConstInfo is a big win.

  • Optimised case trees.

Since we memoise getConstInfo we can do some preprocessing of the definitions, returning a CompactDef instead of a Definition. In particular we streamline the case trees used for matching in a few ways:

  • Drop constructor arity information.
  • Use NameId instead of QName as map keys.
  • Special branch for natural number successor.

None of these changes would make sense to incorporate into the actual case trees. The first two loses information that we need in other places and the third would complicate a lot of code working with case trees.

  • Optimised parallel substitution.

When substituting arguments into function bodies we always have a complete (a term for every free variable) parallel substitution. We run an specialised substitution for this case that falls back to normal substitution when it hits a binder.

Synopsis

Documentation

fastReduce :: Bool -> Term -> ReduceM (Blocked Term) Source #

First argument: allow non-terminating reductions.