Compress paged link-store references

tsc/internal/checker/links.go

@@ -28,7 +28,11 @@ func (s *nodeLinkStore[V]) Get(node *ast.Node) *V {
 }
 
 func (s *nodeLinkStore[V]) Has(node *ast.Node) bool {
-	return s.TryGet(node) != nil
+	id := uint64(s.gen.GetNodeId(node))
+	if id >= ast.BlockIdOffset && id < ast.BlockIdOffset+maxPageLinkCount {
+		return s.pages.Has(id - ast.BlockIdOffset)
+	}
+	return s.links.Has(id)
 }
 
 func (s *nodeLinkStore[V]) TryGet(node *ast.Node) *V {
@@ -54,7 +58,11 @@ func (s *symbolLinkStore[V]) Get(symbol *ast.Symbol) *V {
 }
 
 func (s *symbolLinkStore[V]) Has(symbol *ast.Symbol) bool {
-	return s.TryGet(symbol) != nil
+	id := uint64(s.gen.GetSymbolId(symbol))
+	if id >= ast.BlockIdOffset && id < ast.BlockIdOffset+maxPageLinkCount {
+		return s.pages.Has(id - ast.BlockIdOffset)
+	}
+	return s.links.Has(id)
 }
 
 func (s *symbolLinkStore[V]) TryGet(symbol *ast.Symbol) *V {

tsc/internal/core/arena.go

@@ -1,6 +1,7 @@
 package core
 
 import (
+	"math/bits"
 	"slices"
 	"unsafe"
 )
@@ -69,3 +70,41 @@ func nextArenaSize[T any](size int) int {
 	size = min(size*2, 256, maxSize)
 	return size
 }
+
+// IndexedArena keeps values at stable addresses and references them with uint32
+// indexes instead of Go pointers. Index zero is nil. Unlike Arena, it retains all
+// blocks so that an index alone keeps its value reachable through the arena.
+type IndexedArena[T any] struct {
+	blocks [][]T
+	size   uint32
+}
+
+const indexedArenaPageShift = 8
+
+func (a *IndexedArena[T]) New() (uint32, *T) {
+	if a.size == ^uint32(0) {
+		panic("IndexedArena exhausted")
+	}
+	if len(a.blocks) == 0 || len(a.blocks[len(a.blocks)-1]) == cap(a.blocks[len(a.blocks)-1]) {
+		size := 1 << min(len(a.blocks), indexedArenaPageShift)
+		a.blocks = append(a.blocks, make([]T, 0, size))
+	}
+	block := &a.blocks[len(a.blocks)-1]
+	*block = (*block)[:len(*block)+1]
+	a.size++
+	return a.size, &(*block)[len(*block)-1]
+}
+
+func (a *IndexedArena[T]) Get(index uint32) *T {
+	if index == 0 {
+		return nil
+	}
+	// The initial blocks have sizes 1, 2, 4, ..., 128. Subsequent
+	// blocks have 256 entries, avoiding geometric over-allocation.
+	if index < 1<<indexedArenaPageShift {
+		block := bits.Len32(index) - 1
+		return &a.blocks[block][index-(1<<block)]
+	}
+	index -= 1 << indexedArenaPageShift
+	return &a.blocks[indexedArenaPageShift+int(index>>indexedArenaPageShift)][index&((1<<indexedArenaPageShift)-1)]
+}

tsc/internal/core/linkstore.go

@@ -40,12 +40,12 @@ const (
 )
 
 // PagedLinkStore implements a sparse-array-like structure for storing elements keyed by dense uint64 keys.
-// Elements are allocated in an area, element references are stored in fixed-size pages of 256 entries, and
-// an index of pages is maintained in a growable list.
+// Elements are allocated in an arena, and compressed indexes are stored in fixed-size pages of 256 entries.
+// The pages contain no Go pointers, halving their size on 64-bit hosts and avoiding GC scans of each entry.
 
 type PagedLinkStore[V any] struct {
-	pages []*[LinkPageSize]*V
-	arena Arena[V]
+	pages []*[LinkPageSize]uint32
+	arena IndexedArena[V]
 }
 
 func (s *PagedLinkStore[V]) Get(key uint64) *V {
@@ -56,26 +56,28 @@ func (s *PagedLinkStore[V]) Get(key uint64) *V {
 	}
 	page := s.pages[pageIndex]
 	if page == nil {
-		page = new([LinkPageSize]*V)
+		page = new([LinkPageSize]uint32)
 		s.pages[pageIndex] = page
 	}
-	link := page[key&LinkPageMask]
-	if link == nil {
-		link = s.arena.New()
-		page[key&LinkPageMask] = link
+	index := page[key&LinkPageMask]
+	if index == 0 {
+		id, value := s.arena.New()
+		page[key&LinkPageMask] = id
+		return value
 	}
-	return link
+	return s.arena.Get(index)
 }
 
 func (s *PagedLinkStore[V]) Has(key uint64) bool {
-	return s.TryGet(key) != nil
+	pageIndex := key >> LinkPageShift
+	return int(pageIndex) < len(s.pages) && s.pages[pageIndex] != nil && s.pages[pageIndex][key&LinkPageMask] != 0
 }
 
 func (s *PagedLinkStore[V]) TryGet(key uint64) *V {
 	pageIndex := key >> LinkPageShift
 	if int(pageIndex) < len(s.pages) {
 		if page := s.pages[pageIndex]; page != nil {
-			return page[key&LinkPageMask]
+			return s.arena.Get(page[key&LinkPageMask])
 		}
 	}
 	return nil

tsc/internal/core/linkstore_test.go

@@ -0,0 +1,145 @@
+package core_test
+
+import (
+	"runtime"
+	"strconv"
+	"testing"
+
+	"github.com/microsoft/TypeScript/tsc/internal/core"
+	"gotest.tools/v3/assert"
+)
+
+func TestLinkStore(t *testing.T) {
+	t.Parallel()
+	type value struct {
+		number int
+		text   *string
+	}
+	var store core.LinkStore[int, value]
+	assert.Assert(t, !store.Has(0))
+	assert.Assert(t, store.TryGet(0) == nil)
+
+	values := make([]*value, 4096)
+	for i := range values {
+		v := store.Get(i)
+		assert.Equal(t, *v, value{})
+		v.number = i
+		v.text = new("retained")
+		values[i] = v
+	}
+	runtime.GC()
+	for i, v := range values {
+		assert.Assert(t, store.Has(i))
+		assert.Equal(t, store.Get(i), v)
+		assert.Equal(t, store.TryGet(i), v)
+		assert.Equal(t, v.number, i)
+		assert.Equal(t, *v.text, "retained")
+	}
+	assert.Assert(t, !store.Has(len(values)))
+	assert.Assert(t, store.TryGet(len(values)) == nil)
+}
+
+func TestPagedLinkStore(t *testing.T) {
+	t.Parallel()
+	var store core.PagedLinkStore[int]
+	keys := []uint64{0, 255, 256, 65535, 16777215}
+	values := make([]*int, len(keys))
+	for _, key := range keys {
+		assert.Assert(t, !store.Has(key))
+	}
+	for i, key := range keys {
+		v := store.Get(key)
+		assert.Equal(t, *v, 0)
+		*v = i + 1
+		values[i] = v
+	}
+	runtime.GC()
+	for i, key := range keys {
+		assert.Assert(t, store.Has(key))
+		assert.Equal(t, store.Get(key), values[i])
+		assert.Equal(t, *store.TryGet(key), i+1)
+	}
+	assert.Assert(t, !store.Has(1))
+	assert.Assert(t, store.TryGet(1) == nil)
+}
+
+func BenchmarkLinkStore(b *testing.B) {
+	type value struct {
+		typ   *int
+		flags uint32
+	}
+	b.ReportAllocs()
+	for b.Loop() {
+		var store core.LinkStore[int, value]
+		for i := range 10000 {
+			store.Get(i).flags = uint32(i)
+		}
+		for i := range 10000 {
+			if store.Get(i).flags != uint32(i) {
+				b.Fatal("link changed during arena growth")
+			}
+		}
+
+	}
+}
+
+func TestIndexedArena(t *testing.T) {
+	t.Parallel()
+	type value struct {
+		number int
+		text   *string
+	}
+	var arena core.IndexedArena[value]
+	assert.Assert(t, arena.Get(0) == nil)
+	values := make([]*value, 4096)
+	for i := range values {
+		index, v := arena.New()
+		assert.Equal(t, index, uint32(i+1))
+		assert.Equal(t, *v, value{})
+		v.number = i
+		v.text = new("retained")
+		values[i] = v
+	}
+	runtime.GC()
+	for i, v := range values {
+		assert.Equal(t, arena.Get(uint32(i+1)), v)
+		assert.Equal(t, v.number, i)
+		assert.Equal(t, *v.text, "retained")
+	}
+}
+
+func TestIndexedArenaReachability(t *testing.T) {
+	t.Parallel()
+	var arena core.IndexedArena[*string]
+	for i := range 4096 {
+		_, value := arena.New()
+		*value = new(strconv.Itoa(i))
+	}
+	runtime.GC()
+	for i := range 4096 {
+		assert.Equal(t, **arena.Get(uint32(i + 1)), strconv.Itoa(i))
+	}
+}
+
+func BenchmarkPagedLinkStore(b *testing.B) {
+	for _, stride := range []uint64{1, 16, 256} {
+		b.Run(strconv.FormatUint(stride, 10), func(b *testing.B) {
+			type value struct {
+				typ   *int
+				flags uint32
+			}
+			b.ReportAllocs()
+			for b.Loop() {
+				var store core.PagedLinkStore[value]
+				for i := range 10000 {
+					store.Get(uint64(i) * stride).flags = uint32(i)
+				}
+				for i := range 10000 {
+					if store.Get(uint64(i)*stride).flags != uint32(i) {
+						b.Fatal("link changed during arena growth")
+					}
+				}
+			}
+		})
+	}
+}