hegg
Fast equality saturation in Haskell
Based on egg: Fast and Extensible Equality Saturation, Relational E-matching and the rust implementation.
Equality Saturation and E-graphs
Suggested material on equality saturation and e-graphs for beginners
Equality saturation in Haskell
To get a feel for how we can use hegg
and do equality saturation in Haskell,
we'll write a simple numeric symbolic manipulation library that can simplify expressions
according to a set of rewrite rules by leveraging equality saturation.
If you've never heard of symbolic mathematics you might get some intuition from
reading Let’s Program a Calculus
Student first.
Syntax
We'll start by defining the abstract syntax tree for our simple symbolic expressions:
data SymExpr = Const Double
| Symbol String
| SymExpr :+: SymExpr
| SymExpr :*: SymExpr
| SymExpr :/: SymExpr
infix 6 :+:
infix 7 :*:, :/:
e1 :: SymExpr
e1 = (Symbol "x" :*: Const 2) :/: (Const 2) -- (x*2)/2
You might notice that (x*2)/2
is the same as just x
. Our goal is to get
equality saturation to do that for us.
Our second step is to instance Language
for our SymExpr
Language
Language
is the required constraint on expressions that are to be
represented in e-graph and on which equality saturation can be run:
type Language l = (Traversable l, ∀ a. Ord a => Ord (l a))
To declare a Language
we must write the "base functor" of SymExpr
(i.e. use
a type parameter where the recursion points used to be in the original
SymExpr
), then instance Traversable l
, ∀ a. Ord a => Ord (l a)
(we can do
it automatically through deriving), and write an Analysis
instance for it (see
next section).
data SymExpr a = Const Double
| Symbol String
| a :+: a
| a :*: a
| a :/: a
deriving (Eq, Ord, Show, Functor, Foldable, Traversable)
infix 6 :+:
infix 7 :*:, :/:
Suggested reading on defining recursive data types in their parametrized
version: Introduction To Recursion
Schemes
If we now wanted to represent an expression, we'd write it in its
fixed-point form
e1 :: Fix SymExpr
e1 = Fix (Fix (Fix (Symbol "x") :*: Fix (Const 2)) :/: (Fix (Const 2))) -- (x*2)/2
Then, we define an Analysis
for our SymExpr
.
Analysis
E-class analysis is first described in egg: Fast and Extensible Equality
Saturation as a way to make equality
saturation more extensible.
With it, we can attach analysis data from a semilattice to each e-class. More
can be read about e-class analysis in the Data.Equality.Analsysis
module and
in the paper.
We can easily define constant folding (2+2
being simplified to 4
) through
an Analysis
instance.
An Analysis
is defined over a domain
and a language
. To define constant
folding, we'll say the domain is Maybe Double
to attach a value of that type to
each e-class, where Nothing
indicates the e-class does not currently have a
constant value and Just i
means the e-class has constant value i
.
instance Analysis (Maybe Double) SymExpr
makeA = ...
joinA = ...
modifyA = ...
Let's now understand and implement the three methods of the analysis instance we want.
makeA
is called when a new e-node is added to a new e-class, and constructs
for the new e-class a new value of the domain to be associated with it, always
by accessing the associated data of the node's children data. Its type is l domain -> domain
, so note that the e-node's children associated data is
directly available in place of the actual children.
We want to associate constant data to the e-class, so we must find if the
e-node has a constant value or otherwise return Nothing
:
makeA :: SymExpr (Maybe Double) -> Maybe Double
makeA = \case
Const x -> Just x
Symbol _ -> Nothing
x :+: y -> (+) <$> x <*> y
x :*: y -> (*) <$> x <*> y
x :/: y -> (/) <$> x <*> y
joinA
is called when e-classes c1 c2 are being merged into c. In this case, we
must join the e-class data from both classes to form the e-class data to be
associated with new e-class c. Its type is domain -> domain -> domain
. In our
case, to merge Just _
with Nothing
we simply take the Just
, and if we
merge two e-classes with a constant value (that is, both are Just
), then the
constant value is the same (or something went very wrong) and we just keep it.
joinA :: Maybe Double -> Maybe Double -> Maybe Double
joinA Nothing (Just x) = Just x
joinA (Just x) Nothing = Just x
joinA Nothing Nothing = Nothing
joinA (Just x) (Just y) = if x == y then Just x else error "ouch, that shouldn't have happened"
Finally, modifyA
describes how an e-class should (optionally) be modified
according to the e-class data and what new language expressions are to be added
to the e-class also w.r.t. the e-class data.
Its type is ClassId -> EGraph domain l -> EGraph domain l
, where the first argument
is the id of the class to modify (the class which prompted the modification),
and then receives and returns an e-graph, in which the e-class has been
modified. For our example, if the e-class has a constant value associated to
it, we want to create a new e-class with that constant value and merge it to
this e-class.
-- import Data.Equality.Graph.Lens ((^.), _class, _data)
modifyA :: ClassId -> EGraph (Maybe Double) SymExpr -> EGraph (Maybe Double) SymExpr
modifyA c egr
= case egr ^._class c._data of
Nothing -> egr
Just i ->
let (c', egr') = represent (Fix (Const i)) egr
in snd $ merge c c' egr'
Modify is a bit trickier than the other methods, but it allows our e-graph to
change based on the e-class analysis data. Note that the method is optional and
there's a default implementation for it which doesn't change the e-class or adds
anything to it. Analysis data can be otherwise used, e.g., to inform rewrite
conditions.
By instancing this e-class analysis, all e-classes that have a constant value
associated to them will also have an e-node with a constant value. This is great
for our simple symbolic library because it means if we ever find e.g. an
expression equal to 3+1
, we'll also know it to be equal to 4
, which is a
better result than 3+1
(we've then successfully implemented constant folding).
If, otherwise, we didn't want to use an analysis, we could specify the analysis
domain as ()
which will make the analysis do nothing, because there's an
instance polymorphic over lang
for ()
that looks like this:
instance Analysis () lang where
makeA _ = ()
joinA _ _ = ()
Equality saturation
Equality saturation is defined as the function
equalitySaturation :: forall l. Language l
=> Fix l -- ^ Expression to run equality saturation on
-> [Rewrite l] -- ^ List of rewrite rules
-> CostFunction l -- ^ Cost function to extract the best equivalent representation
-> (Fix l, EGraph l) -- ^ Best equivalent expression and resulting e-graph
To recap, our goal is to reach x
starting from (x*2)/2
by means of equality
saturation.
We already have a starting expression, so we're missing a list of rewrite rules
([Rewrite l]
) and a cost function (CostFunction
).
Cost function
Picking up the easy one first:
type CostFunction l cost = l cost -> cost
A cost function is used to attribute a cost to representations in the e-graph and to extract the best one.
The first type parameter l
is the language we're going to attribute a cost to, and
the second type parameter cost
is the type with which we will model cost. For
the cost function to be valid, cost
must instance Ord
.
We'll say Const
s and Symbol
s are the cheapest and then in increasing cost we
have :+:
, :*:
and :/:
, and model cost with the Int
type.
cost :: CostFunction SymExpr Int
cost = \case
Const x -> 1
Symbol x -> 1
c1 :+: c2 -> c1 + c2 + 2
c1 :*: c2 -> c1 + c2 + 3
c1 :/: c2 -> c1 + c2 + 4
Rewrite rules
Rewrite rules are transformations applied to matching expressions represented in
an e-graph.
We can write simple rewrite rules and conditional rewrite rules, but we'll only look at the simple ones.
A simple rewrite is formed of its left hand side and right hand side. When the
left hand side is matched in the e-graph, the right hand side is added to the
e-class where the left hand side was found.
data Rewrite lang = Pattern lang := Pattern lang -- Simple rewrite rule
| Rewrite lang :| RewriteCondition lang -- Conditional rewrite rule
A Pattern
is basically an expression that might contain variables and which can be matched against actual expressions.
data Pattern lang
= NonVariablePattern (lang (Pattern lang))
| VariablePattern Var
A patterns is defined by its non-variable and variable parts, and can be
constructed directly or using the helper function pat
and using
OverloadedStrings
for the variables, where pat
is just a synonym for
NonVariablePattern
and a string literal "abc"
is turned into a Pattern
constructed with VariablePattern
.
We can then write the following very specific set of rewrite rules to simplify
our simple symbolic expressions.
rewrites :: [Rewrite SymExpr]
rewrites =
[ pat (pat ("a" :*: "b") :/: "c") := pat ("a" :*: pat ("b" :/: "c"))
, pat ("x" :/: "x") := pat (Const 1)
, pat ("x" :*: (pat (Const 1))) := "x"
]
Equality saturation, again
We can now run equality saturation on our expression!
let expr = fst (equalitySaturation e1 rewrites cost)
And upon printing we'd see expr = Symbol "x"
!
If we had instead e2 = Fix (Fix (Fix (Symbol "x") :/: Fix (Symbol "x")) :+: (Fix (Const 3))) -- (x/x)+3
, we'd get expr = Const 4
because of our rewrite
rules put together with our constant folding!
This was a first introduction which skipped over some details but that tried to
walk through fundamental concepts for using e-graphs and equality saturation
with this library.
The final code for this tutorial is available under test/SimpleSym.hs
A more complicated symbolic rewrite system which simplifies some derivatives and
integrals was written for the testsuite. It can be found at test/Sym.hs
.
This library could also be used not only for equality-saturation but also for
the equality-graphs and other equality-things (such as e-matching) available.
For example, using just the e-graphs from Data.Equality.Graph
to improve GHC's
pattern match checker (https://gitlab.haskell.org/ghc/ghc/-/issues/19272).
Profiling
Notes on profiling for development.
For producing the info table, ghc-options must include -finfo-table-map -fdistinct-constructor-tables
cabal run --enable-profiling hegg-test -- +RTS -p -s -hi -l-agu
ghc-prof-flamegraph hegg-test.prof
eventlog2html hegg-test.eventlog
open hegg-test.svg
open hegg-test.eventlog.html