luau/EqSat/include/Luau/EGraph.h
Alexander McCord e1bf6289c7
Equality graphs (#1285)
Working towards a full e-graph implementation as described by the [egg
paper](https://arxiv.org/pdf/2004.03082).

The type system has a couple of places where e-graphs would've been
useful and solved some classes of problems trivially. For example:

1. Normalization and simplification cannot handle cyclic types due to
the nature of their implementation.
2. Normalization can't tell when two tables or functions are equivalent,
but simplification theoretically can albeit not implemented.
3. Normalization requires deep normalization for inhabitance check,
whereas simplification would've returned the `never` type itself
indicating uninhabited.
4. Simplification requires constraint ordering to have perfect timing to
simplify.
5. Adding a rewrite rule requires implementing it twice, once in
simplification and once again in normalization with completely different
code design making it hard to verify that their behavior is materially
equivalent.
6. In cases where we must cache for performance, two different types
that are isomorphic have different cache entries resulting in cache
misses.
7. Type family reduction can handle cyclic type families, but only if
the cycle is not obscured by a different type family instance. (`t1
where t1 = union<number, add<t1, number>>` is irreducible)

I think we're getting the point!

---

Currently the implementation is missing a few features that makes
e-graphs actually useful. Those will be coming in a future PR.

1. Pattern matching,
6. Applying rewrites,
7. Rewrite until saturation, and
8. Extracting the best e-node according to some cost function.
2024-07-16 10:35:20 -07:00

229 lines
6.1 KiB
C++
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

// This file is part of the Luau programming language and is licensed under MIT License; see LICENSE.txt for details
#pragma once
#include "Luau/Common.h"
#include "Luau/Id.h"
#include "Luau/Language.h"
#include "Luau/UnionFind.h"
#include "Luau/VecDeque.h"
#include <optional>
#include <unordered_map>
#include <vector>
namespace Luau::EqSat
{
template<typename L, typename N>
struct EGraph;
template<typename L, typename N>
struct Analysis final
{
N analysis;
using D = typename N::Data;
template<typename T>
static D fnMake(const N& analysis, const EGraph<L, N>& egraph, const L& enode)
{
return analysis.make(egraph, *enode.template get<T>());
}
template<typename... Ts>
D make(const EGraph<L, N>& egraph, const Language<Ts...>& enode) const
{
using FnMake = D (*)(const N&, const EGraph<L, N>&, const L&);
static constexpr FnMake tableMake[sizeof...(Ts)] = {&fnMake<Ts>...};
return tableMake[enode.index()](analysis, egraph, enode);
}
void join(D& a, const D& b) const
{
return analysis.join(a, b);
}
};
/// Each e-class is a set of e-nodes representing equivalent terms from a given language,
/// and an e-node is a function symbol paired with a list of children e-classes.
template<typename L, typename D>
struct EClass final
{
Id id;
std::vector<L> nodes;
D data;
std::vector<std::pair<L, Id>> parents;
};
/// See <https://arxiv.org/pdf/2004.03082>.
template<typename L, typename N>
struct EGraph final
{
Id find(Id id) const
{
return unionfind.find(id);
}
std::optional<Id> lookup(const L& enode) const
{
LUAU_ASSERT(isCanonical(enode));
if (auto it = hashcons.find(enode); it != hashcons.end())
return it->second;
return std::nullopt;
}
Id add(L enode)
{
canonicalize(enode);
if (auto id = lookup(enode))
return *id;
Id id = makeEClass(enode);
return id;
}
void merge(Id id1, Id id2)
{
id1 = find(id1);
id2 = find(id2);
if (id1 == id2)
return;
unionfind.merge(id1, id2);
EClass<L, typename N::Data>& eclass1 = get(id1);
EClass<L, typename N::Data> eclass2 = std::move(get(id2));
classes.erase(id2);
worklist.reserve(worklist.size() + eclass2.parents.size());
for (auto [enode, id] : eclass2.parents)
worklist.push_back({std::move(enode), id});
analysis.join(eclass1.data, eclass2.data);
}
void rebuild()
{
while (!worklist.empty())
{
auto [enode, id] = worklist.back();
worklist.pop_back();
repair(get(find(id)));
}
}
size_t size() const
{
return classes.size();
}
EClass<L, typename N::Data>& operator[](Id id)
{
return get(find(id));
}
const EClass<L, typename N::Data>& operator[](Id id) const
{
return const_cast<EGraph*>(this)->get(find(id));
}
private:
Analysis<L, N> analysis;
/// A union-find data structure 𝑈 stores an equivalence relation over e-class ids.
UnionFind unionfind;
/// The e-class map 𝑀 maps e-class ids to e-classes. All equivalent e-class ids map to the same
/// e-class, i.e., 𝑎 ≡id 𝑏 iff 𝑀[𝑎] is the same set as 𝑀[𝑏]. An e-class id 𝑎 is said to refer to the
/// e-class 𝑀[find(𝑎)].
std::unordered_map<Id, EClass<L, typename N::Data>> classes;
/// The hashcons 𝐻 is a map from e-nodes to e-class ids.
std::unordered_map<L, Id, typename L::Hash> hashcons;
VecDeque<std::pair<L, Id>> worklist;
private:
void canonicalize(L& enode)
{
// An e-node 𝑛 is canonical iff 𝑛 = canonicalize(𝑛), where
// canonicalize(𝑓(𝑎1, 𝑎2, ...)) = 𝑓(find(𝑎1), find(𝑎2), ...).
for (Id& id : enode.operands())
id = find(id);
}
bool isCanonical(const L& enode) const
{
bool canonical = true;
for (Id id : enode.operands())
canonical &= (id == find(id));
return canonical;
}
Id makeEClass(const L& enode)
{
LUAU_ASSERT(isCanonical(enode));
Id id = unionfind.makeSet();
classes.insert_or_assign(id, EClass<L, typename N::Data>{
id,
{enode},
analysis.make(*this, enode),
{},
});
for (Id operand : enode.operands())
get(operand).parents.push_back({enode, id});
worklist.push_back({enode, id});
hashcons.insert_or_assign(enode, id);
return id;
}
// Looks up for an eclass from a given non-canonicalized `id`.
// For a canonicalized eclass, use `get(find(id))` or `egraph[id]`.
EClass<L, typename N::Data>& get(Id id)
{
return classes.at(id);
}
void repair(EClass<L, typename N::Data>& eclass)
{
// In the egg paper, the `repair` function makes use of two loops over the `eclass.parents`
// by first erasing the old enode entry, and adding back the canonicalized enode with the canonical id.
// And then in another loop that follows, deduplicate it.
//
// Here, we unify the two loops. I think it's equivalent?
// After canonicalizing the enodes, the eclass may contain multiple enodes that are equivalent.
std::unordered_map<L, Id, typename L::Hash> map;
for (auto& [enode, id] : eclass.parents)
{
// By removing the old enode from the hashcons map, we will always find our new canonicalized eclass id.
hashcons.erase(enode);
canonicalize(enode);
hashcons.insert_or_assign(enode, find(id));
if (auto it = map.find(enode); it != map.end())
merge(id, it->second);
map.insert_or_assign(enode, find(id));
}
eclass.parents.clear();
for (auto it = map.begin(); it != map.end();)
{
auto node = map.extract(it++);
eclass.parents.emplace_back(std::move(node.key()), node.mapped());
}
}
};
} // namespace Luau::EqSat