From 8da430f4c95f8e31aef00863a34531377e23f814 Mon Sep 17 00:00:00 2001
From: Jacob Lifshay <programmerjake@gmail.com>
Date: Tue, 3 Feb 2026 17:30:27 -0800
Subject: [PATCH] speed up LazyInterned by redoing caching using RwLock and add
 a thread-local cache

---
 crates/fayalite/src/array.rs           |  32 +++-
 crates/fayalite/src/intern.rs          | 244 +++++++++++++++----------
 crates/fayalite/src/intern/type_map.rs |   4 +-
 crates/fayalite/src/phantom_const.rs   |  16 +-
 4 files changed, 189 insertions(+), 107 deletions(-)

diff --git a/crates/fayalite/src/array.rs b/crates/fayalite/src/array.rs
index 4e2b223..6ca6809 100644
--- a/crates/fayalite/src/array.rs
+++ b/crates/fayalite/src/array.rs
@@ -109,14 +109,42 @@ impl<T: StaticType, Len: KnownSize> Default for ArrayType<T, Len> {
     }
 }
 
+struct MakeType<T: StaticType>(Interned<T>);
+
+impl<T: StaticType> From<MakeType<T>> for Interned<T> {
+    fn from(value: MakeType<T>) -> Self {
+        value.0
+    }
+}
+
+impl<T: StaticType> Default for MakeType<T> {
+    fn default() -> Self {
+        Self(T::TYPE.intern_sized())
+    }
+}
+
+struct MakeMaskType<T: StaticType>(Interned<T::MaskType>);
+
+impl<T: StaticType> From<MakeMaskType<T>> for Interned<T::MaskType> {
+    fn from(value: MakeMaskType<T>) -> Self {
+        value.0
+    }
+}
+
+impl<T: StaticType> Default for MakeMaskType<T> {
+    fn default() -> Self {
+        Self(T::MASK_TYPE.intern_sized())
+    }
+}
+
 impl<T: StaticType, Len: KnownSize> StaticType for ArrayType<T, Len> {
     const TYPE: Self = Self {
-        element: LazyInterned::new_lazy(&|| T::TYPE.intern_sized()),
+        element: LazyInterned::new_const::<MakeType<T>>(),
         len: Len::SIZE,
         type_properties: Self::TYPE_PROPERTIES,
     };
     const MASK_TYPE: Self::MaskType = ArrayType::<T::MaskType, Len> {
-        element: LazyInterned::new_lazy(&|| T::MASK_TYPE.intern_sized()),
+        element: LazyInterned::new_const::<MakeMaskType<T>>(),
         len: Len::SIZE,
         type_properties: Self::MASK_TYPE_PROPERTIES,
     };
diff --git a/crates/fayalite/src/intern.rs b/crates/fayalite/src/intern.rs
index 9ffd251..b78aa59 100644
--- a/crates/fayalite/src/intern.rs
+++ b/crates/fayalite/src/intern.rs
@@ -9,12 +9,12 @@ use serde::{Deserialize, Serialize};
 use std::{
     any::{Any, TypeId},
     borrow::{Borrow, Cow},
+    cell::RefCell,
     cmp::Ordering,
     ffi::{OsStr, OsString},
     fmt,
     hash::{BuildHasher, Hash, Hasher},
     iter::FusedIterator,
-    marker::PhantomData,
     ops::Deref,
     path::{Path, PathBuf},
     sync::RwLock,
@@ -23,51 +23,172 @@ use std::{
 mod interner;
 mod type_map;
 
-pub trait LazyInternedTrait<T: ?Sized + Send + Sync + 'static>: Send + Sync + Any {
-    fn get(&self) -> Interned<T>;
+/// invariant: T must be zero-sized, `type_id` is unique for every possible T value.
+struct LazyInternedLazyInner<T: ?Sized + 'static> {
+    type_id: TypeId,
+    value: T,
 }
 
-impl<T: ?Sized + Send + Sync + 'static, F: ?Sized + Fn() -> Interned<T> + Send + Sync + Any>
-    LazyInternedTrait<T> for F
-{
-    fn get(&self) -> Interned<T> {
-        self()
+impl<T: ?Sized + 'static> Hash for LazyInternedLazyInner<T> {
+    fn hash<H: Hasher>(&self, state: &mut H) {
+        let Self { type_id, value: _ } = self;
+        type_id.hash(state);
     }
 }
 
-#[repr(transparent)]
-pub struct LazyInternedFn<T: ?Sized + Send + Sync + 'static>(pub &'static dyn LazyInternedTrait<T>);
+impl<T: ?Sized + 'static> PartialEq for LazyInternedLazyInner<T> {
+    fn eq(&self, other: &Self) -> bool {
+        let Self { type_id, value: _ } = self;
+        *type_id == other.type_id
+    }
+}
 
-impl<T: ?Sized + Send + Sync + 'static> Copy for LazyInternedFn<T> {}
+impl<T: ?Sized + 'static> Eq for LazyInternedLazyInner<T> {}
 
-impl<T: ?Sized + Send + Sync + 'static> Clone for LazyInternedFn<T> {
+impl<T: ?Sized + 'static> fmt::Debug for LazyInternedLazyInner<T> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("LazyInternedLazyInner")
+            .finish_non_exhaustive()
+    }
+}
+
+impl<T: ?Sized + 'static> LazyInternedLazyInner<T> {
+    const fn new(value: T) -> Self
+    where
+        T: Sized,
+    {
+        const { assert!(size_of::<T>() == 0) };
+        Self {
+            type_id: TypeId::of::<T>(),
+            value,
+        }
+    }
+}
+
+pub struct LazyInternedLazy<T: ?Sized + Send + Sync + 'static>(
+    &'static LazyInternedLazyInner<dyn Fn() -> Interned<T> + Send + Sync>,
+);
+
+impl<T: ?Sized + Send + Sync + 'static> LazyInternedLazy<T> {
+    pub const fn new_const<V: Default + Into<Interned<T>>>() -> Self {
+        Self(&const { LazyInternedLazyInner::new(|| V::default().into()) })
+    }
+    pub const fn new_const_default() -> Self
+    where
+        Interned<T>: Default,
+    {
+        Self::new_const::<Interned<T>>()
+    }
+    pub fn interned(self) -> Interned<T> {
+        struct Map(hashbrown::HashTable<(TypeId, &'static (dyn Any + Send + Sync))>);
+        impl Map {
+            const EMPTY: Self = Self(hashbrown::HashTable::new());
+            fn get<T: ?Sized + Send + Sync + 'static>(
+                &self,
+                lazy_interned_lazy: LazyInternedLazy<T>,
+                hash: u64,
+            ) -> Option<&'static Interned<T>> {
+                let &(_, v) = self.0.find(hash, |v| v.0 == lazy_interned_lazy.0.type_id)?;
+                let Some(retval) = v.downcast_ref::<Interned<T>>() else {
+                    unreachable!();
+                };
+                Some(retval)
+            }
+            fn get_or_insert<T: ?Sized + Send + Sync + 'static>(
+                &mut self,
+                lazy_interned_lazy: LazyInternedLazy<T>,
+                hash: u64,
+                v: &'static Interned<T>,
+            ) -> &'static Interned<T> {
+                let entry = self
+                    .0
+                    .entry(
+                        hash,
+                        |&(k, _)| k == lazy_interned_lazy.0.type_id,
+                        |&(k, _)| type_map::TypeIdBuildHasher.hash_one(k),
+                    )
+                    .or_insert_with(|| (lazy_interned_lazy.0.type_id, v));
+                let &(_, v) = entry.get();
+                let Some(retval) = v.downcast_ref::<Interned<T>>() else {
+                    unreachable!();
+                };
+                retval
+            }
+        }
+        static GLOBAL_CACHE: RwLock<Map> = RwLock::new(Map::EMPTY);
+        #[cold]
+        fn insert_in_thread_local_cache<T: ?Sized + Send + Sync + 'static>(
+            cache: &RefCell<Map>,
+            this: LazyInternedLazy<T>,
+            hash: u64,
+        ) -> Interned<T> {
+            let read_lock = GLOBAL_CACHE.read().unwrap();
+            let v = read_lock.get(this, hash);
+            drop(read_lock);
+            let v = v.unwrap_or_else(|| {
+                let v = Box::leak(Box::new((this.0.value)()));
+                GLOBAL_CACHE.write().unwrap().get_or_insert(this, hash, v)
+            });
+            *cache.borrow_mut().get_or_insert(this, hash, v)
+        }
+        thread_local! {
+            static THREAD_LOCAL_CACHE: RefCell<Map> = const { RefCell::new(Map::EMPTY) };
+        }
+        let hash = type_map::TypeIdBuildHasher.hash_one(self.0.type_id);
+        THREAD_LOCAL_CACHE.with(|cache| {
+            let borrow = cache.borrow();
+            if let Some(v) = borrow.get(self, hash) {
+                *v
+            } else {
+                drop(borrow);
+                insert_in_thread_local_cache(cache, self, hash)
+            }
+        })
+    }
+}
+
+impl<T: ?Sized + Send + Sync + 'static> Copy for LazyInternedLazy<T> {}
+
+impl<T: ?Sized + Send + Sync + 'static> Clone for LazyInternedLazy<T> {
     fn clone(&self) -> Self {
         *self
     }
 }
 
-impl<T: ?Sized + Send + Sync + 'static> Hash for LazyInternedFn<T> {
+impl<T: ?Sized + Send + Sync + 'static> Hash for LazyInternedLazy<T> {
     fn hash<H: Hasher>(&self, state: &mut H) {
-        self.0.get_ptr_eq_with_type_id().hash(state);
+        self.0.hash(state);
     }
 }
 
-impl<T: ?Sized + Send + Sync + 'static> Eq for LazyInternedFn<T> {}
+impl<T: ?Sized + Send + Sync + 'static> Eq for LazyInternedLazy<T> {}
 
-impl<T: ?Sized + Send + Sync + 'static> PartialEq for LazyInternedFn<T> {
+impl<T: ?Sized + Send + Sync + 'static> PartialEq for LazyInternedLazy<T> {
     fn eq(&self, other: &Self) -> bool {
-        self.0.get_ptr_eq_with_type_id() == other.0.get_ptr_eq_with_type_id()
+        self.0 == other.0
     }
 }
 
 pub enum LazyInterned<T: ?Sized + Send + Sync + 'static> {
     Interned(Interned<T>),
-    Lazy(LazyInternedFn<T>),
+    Lazy(LazyInternedLazy<T>),
 }
 
 impl<T: ?Sized + Send + Sync + 'static> LazyInterned<T> {
-    pub const fn new_lazy(v: &'static dyn LazyInternedTrait<T>) -> Self {
-        Self::Lazy(LazyInternedFn(v))
+    pub const fn new_const<V: Default + Into<Interned<T>>>() -> Self {
+        Self::Lazy(LazyInternedLazy::new_const::<V>())
+    }
+    pub const fn new_const_default() -> Self
+    where
+        Interned<T>: Default,
+    {
+        Self::new_const::<Interned<T>>()
+    }
+    pub fn interned(self) -> Interned<T> {
+        match self {
+            Self::Interned(retval) => retval,
+            Self::Lazy(retval) => retval.interned(),
+        }
     }
 }
 
@@ -79,7 +200,7 @@ impl<T: ?Sized + Sync + Send + 'static> Clone for LazyInterned<T> {
 
 impl<T: ?Sized + Sync + Send + 'static> Copy for LazyInterned<T> {}
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> Deref for LazyInterned<T> {
+impl<T: ?Sized + Sync + Send + 'static> Deref for LazyInterned<T> {
     type Target = T;
 
     fn deref(&self) -> &Self::Target {
@@ -87,9 +208,9 @@ impl<T: ?Sized + Sync + Send + 'static + Intern> Deref for LazyInterned<T> {
     }
 }
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> Eq for LazyInterned<T> where Interned<T>: Eq {}
+impl<T: ?Sized + Sync + Send + 'static> Eq for LazyInterned<T> where Interned<T>: Eq {}
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> PartialEq for LazyInterned<T>
+impl<T: ?Sized + Sync + Send + 'static> PartialEq for LazyInterned<T>
 where
     Interned<T>: PartialEq,
 {
@@ -98,7 +219,7 @@ where
     }
 }
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> Ord for LazyInterned<T>
+impl<T: ?Sized + Sync + Send + 'static> Ord for LazyInterned<T>
 where
     Interned<T>: Ord,
 {
@@ -107,7 +228,7 @@ where
     }
 }
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> PartialOrd for LazyInterned<T>
+impl<T: ?Sized + Sync + Send + 'static> PartialOrd for LazyInterned<T>
 where
     Interned<T>: PartialOrd,
 {
@@ -116,7 +237,7 @@ where
     }
 }
 
-impl<T: ?Sized + Sync + Send + 'static + Intern> Hash for LazyInterned<T>
+impl<T: ?Sized + Sync + Send + 'static> Hash for LazyInterned<T>
 where
     Interned<T>: Hash,
 {
@@ -125,77 +246,6 @@ where
     }
 }
 
-impl<T: ?Sized + Sync + Send + 'static> LazyInterned<T> {
-    pub fn interned(self) -> Interned<T>
-    where
-        T: Intern,
-    {
-        struct MemoizeInterned<T: ?Sized + Intern>(PhantomData<T>);
-
-        impl<T: ?Sized + Intern> Hash for MemoizeInterned<T> {
-            fn hash<H: Hasher>(&self, _state: &mut H) {}
-        }
-
-        impl<T: ?Sized + Intern> PartialEq for MemoizeInterned<T> {
-            fn eq(&self, _other: &Self) -> bool {
-                true
-            }
-        }
-
-        impl<T: ?Sized + Intern> Eq for MemoizeInterned<T> {}
-
-        impl<T: ?Sized + Intern> Clone for MemoizeInterned<T> {
-            fn clone(&self) -> Self {
-                *self
-            }
-        }
-
-        impl<T: ?Sized + Intern> Copy for MemoizeInterned<T> {}
-
-        impl<T: ?Sized + Intern> MemoizeGeneric for MemoizeInterned<T> {
-            type InputRef<'a> = LazyInternedFn<T>;
-
-            type InputOwned = LazyInternedFn<T>;
-
-            type InputCow<'a> = LazyInternedFn<T>;
-
-            type Output = Interned<T>;
-
-            fn input_eq(a: Self::InputRef<'_>, b: Self::InputRef<'_>) -> bool {
-                a == b
-            }
-
-            fn input_borrow(input: &Self::InputOwned) -> Self::InputRef<'_> {
-                *input
-            }
-
-            fn input_cow_into_owned(input: Self::InputCow<'_>) -> Self::InputOwned {
-                input
-            }
-
-            fn input_cow_borrow<'a>(input: &'a Self::InputCow<'_>) -> Self::InputRef<'a> {
-                *input
-            }
-
-            fn input_cow_from_owned<'a>(input: Self::InputOwned) -> Self::InputCow<'a> {
-                input
-            }
-
-            fn input_cow_from_ref(input: Self::InputRef<'_>) -> Self::InputCow<'_> {
-                input
-            }
-
-            fn inner(self, input: Self::InputRef<'_>) -> Self::Output {
-                input.0.get()
-            }
-        }
-        match self {
-            Self::Interned(retval) => retval,
-            Self::Lazy(retval) => MemoizeInterned(PhantomData).get(retval),
-        }
-    }
-}
-
 pub trait InternedCompare {
     type InternedCompareKey: Ord + Hash;
     fn interned_compare_key_ref(this: &Self) -> Self::InternedCompareKey;
diff --git a/crates/fayalite/src/intern/type_map.rs b/crates/fayalite/src/intern/type_map.rs
index 03b68e3..e31a5bf 100644
--- a/crates/fayalite/src/intern/type_map.rs
+++ b/crates/fayalite/src/intern/type_map.rs
@@ -6,7 +6,7 @@ use std::{
     sync::RwLock,
 };
 
-struct TypeIdHasher(u64);
+pub(crate) struct TypeIdHasher(u64);
 
 // assumes TypeId has at least 64 bits that is a good hash
 impl Hasher for TypeIdHasher {
@@ -63,7 +63,7 @@ impl Hasher for TypeIdHasher {
     }
 }
 
-struct TypeIdBuildHasher;
+pub(crate) struct TypeIdBuildHasher;
 
 impl BuildHasher for TypeIdBuildHasher {
     type Hasher = TypeIdHasher;
diff --git a/crates/fayalite/src/phantom_const.rs b/crates/fayalite/src/phantom_const.rs
index 32e9d6b..fb7be6f 100644
--- a/crates/fayalite/src/phantom_const.rs
+++ b/crates/fayalite/src/phantom_const.rs
@@ -4,7 +4,7 @@
 use crate::{
     expr::{Expr, HdlPartialEqImpl, HdlPartialOrdImpl, ToExpr, ValueType},
     int::Bool,
-    intern::{Intern, Interned, InternedCompare, LazyInterned, LazyInternedTrait, Memoize},
+    intern::{Intern, Interned, InternedCompare, LazyInterned, Memoize},
     sim::value::{SimValue, ToSimValue, ToSimValueWithType},
     source_location::SourceLocation,
     ty::{
@@ -240,11 +240,17 @@ impl<T: ?Sized + PhantomConstValue> PhantomConst<T> {
     {
         Self::new_interned(value.intern_deref())
     }
-    pub const fn new_lazy(v: &'static dyn LazyInternedTrait<T>) -> Self {
+    pub const fn new_const<V: Default + Into<Interned<T>>>() -> Self {
         Self {
-            value: LazyInterned::new_lazy(v),
+            value: LazyInterned::new_const::<V>(),
         }
     }
+    pub const fn new_const_default() -> Self
+    where
+        Interned<T>: Default,
+    {
+        Self::new_const::<Interned<T>>()
+    }
     pub fn get(self) -> Interned<T> {
         self.value.interned()
     }
@@ -334,9 +340,7 @@ impl<T: ?Sized + PhantomConstValue> StaticType for PhantomConst<T>
 where
     Interned<T>: Default,
 {
-    const TYPE: Self = PhantomConst {
-        value: LazyInterned::new_lazy(&Interned::<T>::default),
-    };
+    const TYPE: Self = Self::new_const_default();
     const MASK_TYPE: Self::MaskType = ();
     const TYPE_PROPERTIES: TypeProperties = <()>::TYPE_PROPERTIES;
     const MASK_TYPE_PROPERTIES: TypeProperties = <()>::TYPE_PROPERTIES;