cppgc: young-gen: Pack AgeTable knowing allocation granularity

This reduces card granularity from 4096 to 512 bytes with the goal to
improve write barrier filtering.

Bug: chromium:1029379
Change-Id: I22e2a9c61ef4c36c3db65404370213d0a8048e08
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3582393Reviewed-by: Michael Lippautz <mlippautz@chromium.org>
Commit-Queue: Anton Bikineev <bikineev@chromium.org>
Cr-Commit-Position: refs/heads/main@{#79960}

include/cppgc/internal/api-constants.h

@@ -44,6 +44,9 @@ static constexpr size_t kDefaultAlignment = sizeof(void*);
 // Maximum support alignment for a type as in `alignof(T)`.
 static constexpr size_t kMaxSupportedAlignment = 2 * kDefaultAlignment;
 
+// Granularity of heap allocations.
+constexpr size_t kAllocationGranularity = sizeof(void*);
+
 }  // namespace api_constants
 
 }  // namespace internal

include/cppgc/internal/caged-heap-local-data.h

@@ -6,6 +6,8 @@
 #define INCLUDE_CPPGC_INTERNAL_CAGED_HEAP_LOCAL_DATA_H_
 
 #include <array>
+#include <cstddef>
+#include <cstdint>
 
 #include "cppgc/internal/api-constants.h"
 #include "cppgc/internal/logging.h"
@@ -19,32 +21,41 @@ class HeapBase;
 
 #if defined(CPPGC_YOUNG_GENERATION)
 
-// AgeTable contains entries that correspond to 4KB memory regions. Each entry
-// can be in one of three states: kOld, kYoung or kUnknown.
+// AgeTable is the bytemap needed for the fast generation check in the write
+// barrier. AgeTable contains entries that correspond to 512 bytes memory
+// regions (cards). Each entry in the table represents generation of the objects
+// that reside on the corresponding card (young, old or mixed).
 class AgeTable final {
-  static constexpr size_t kGranularityBits = 12;  // 4KiB per byte.
+  static constexpr size_t kRequiredSize = 1 * api_constants::kMB;
+  static constexpr size_t kAllocationGranularity =
+      api_constants::kAllocationGranularity;
 
  public:
-  enum class Age : uint8_t { kOld, kYoung, kUnknown };
+  enum class Age : uint8_t { kOld, kYoung, kMixed };
 
-  static constexpr size_t kEntrySizeInBytes = 1 << kGranularityBits;
+  static constexpr size_t kCardSizeInBytes =
+      (api_constants::kCagedHeapReservationSize / kAllocationGranularity) /
+      kRequiredSize;
 
-  Age& operator[](uintptr_t offset) { return table_[entry(offset)]; }
-  Age operator[](uintptr_t offset) const { return table_[entry(offset)]; }
+  void SetAge(uintptr_t cage_offset, Age age) {
+    table_[card(cage_offset)] = age;
+  }
+  V8_INLINE Age GetAge(uintptr_t cage_offset) const {
+    return table_[card(cage_offset)];
+  }
 
   void Reset(PageAllocator* allocator);
 
  private:
-  static constexpr size_t kAgeTableSize =
-      api_constants::kCagedHeapReservationSize >> kGranularityBits;
-
-  size_t entry(uintptr_t offset) const {
+  V8_INLINE size_t card(uintptr_t offset) const {
+    constexpr size_t kGranularityBits =
+        __builtin_ctz(static_cast<uint32_t>(kCardSizeInBytes));
     const size_t entry = offset >> kGranularityBits;
     CPPGC_DCHECK(table_.size() > entry);
     return entry;
   }
 
-  std::array<Age, kAgeTableSize> table_;
+  std::array<Age, kRequiredSize> table_;
 };
 
 static_assert(sizeof(AgeTable) == 1 * api_constants::kMB,

include/cppgc/internal/write-barrier.h

@@ -405,7 +405,8 @@ void WriteBarrier::GenerationalBarrier(const Params& params, const void* slot) {
   const AgeTable& age_table = local_data.age_table;
 
   // Bail out if the slot is in young generation.
-  if (V8_LIKELY(age_table[params.slot_offset] == AgeTable::Age::kYoung)) return;
+  if (V8_LIKELY(age_table.GetAge(params.slot_offset) == AgeTable::Age::kYoung))
+    return;
 
   GenerationalBarrierSlow(local_data, age_table, slot, params.value_offset);
 }
@@ -420,7 +421,8 @@ void WriteBarrier::GenerationalBarrierForSourceObject(
 
   // Assume that if the first element is in young generation, the whole range is
   // in young generation.
-  if (V8_LIKELY(age_table[params.slot_offset] == AgeTable::Age::kYoung)) return;
+  if (V8_LIKELY(age_table.GetAge(params.slot_offset) == AgeTable::Age::kYoung))
+    return;
 
   GenerationalBarrierForSourceObjectSlow(local_data, inner_pointer);
 }

src/heap/cppgc/marking-verifier.cc

@@ -106,14 +106,12 @@ bool MarkingVerifierBase::VisitHeapObjectHeader(HeapObjectHeader& header) {
 
 #if defined(CPPGC_YOUNG_GENERATION)
   if (collection_type_ == Heap::Config::CollectionType::kMinor) {
-    const auto age = heap_.caged_heap()
-                         .local_data()
-                         .age_table[heap_.caged_heap().OffsetFromAddress(
-                             header.ObjectStart())];
+    const auto age = heap_.caged_heap().local_data().age_table.GetAge(
+        heap_.caged_heap().OffsetFromAddress(header.ObjectStart()));
     if (age == AgeTable::Age::kOld) {
       // Do not verify old objects.
       return true;
-    } else if (age == AgeTable::Age::kUnknown) {
+    } else if (age == AgeTable::Age::kMixed) {
       // If the age is not known, the marked bytes may not be exact as possibly
       // old objects are verified as well.
       verifier_found_marked_bytes_are_exact_ = false;

src/heap/cppgc/object-allocator.cc

@@ -30,7 +30,7 @@ void MarkRangeAsYoung(BasePage* page, Address begin, Address end) {
 #if defined(CPPGC_YOUNG_GENERATION)
   DCHECK_LT(begin, end);
 
-  static constexpr auto kEntrySize = AgeTable::kEntrySizeInBytes;
+  static constexpr auto kEntrySize = AgeTable::kCardSizeInBytes;
 
   const uintptr_t offset_begin = CagedHeap::OffsetFromAddress(begin);
   const uintptr_t offset_end = CagedHeap::OffsetFromAddress(end);
@@ -44,16 +44,16 @@ void MarkRangeAsYoung(BasePage* page, Address begin, Address end) {
 
   auto& age_table = page->heap().caged_heap().local_data().age_table;
   for (auto offset = young_offset_begin; offset < young_offset_end;
-       offset += AgeTable::kEntrySizeInBytes) {
-    age_table[offset] = AgeTable::Age::kYoung;
+       offset += AgeTable::kCardSizeInBytes) {
+    age_table.SetAge(offset, AgeTable::Age::kYoung);
   }
 
   // Set to kUnknown the first and the last regions of the newly allocated
   // linear buffer.
   if (begin != page->PayloadStart() && !IsAligned(offset_begin, kEntrySize))
-    age_table[offset_begin] = AgeTable::Age::kUnknown;
+    age_table.SetAge(offset_begin, AgeTable::Age::kMixed);
   if (end != page->PayloadEnd() && !IsAligned(offset_end, kEntrySize))
-    age_table[offset_end] = AgeTable::Age::kUnknown;
+    age_table.SetAge(offset_end, AgeTable::Age::kMixed);
 #endif
 }
 

src/heap/cppgc/object-allocator.h

@@ -146,6 +146,8 @@ void* ObjectAllocator::AllocateObjectOnSpace(NormalPageSpace& space,
   STATIC_ASSERT(2 * kAllocationGranularity ==
                 api_constants::kMaxSupportedAlignment);
   STATIC_ASSERT(kAllocationGranularity == sizeof(HeapObjectHeader));
+  STATIC_ASSERT(kAllocationGranularity ==
+                api_constants::kAllocationGranularity);
   DCHECK_EQ(2 * sizeof(HeapObjectHeader), static_cast<size_t>(alignment));
   constexpr size_t kAlignment = 2 * kAllocationGranularity;
   constexpr size_t kAlignmentMask = kAlignment - 1;

src/heap/cppgc/write-barrier.cc

@@ -135,7 +135,7 @@ void WriteBarrier::GenerationalBarrierSlow(const CagedHeapLocalData& local_data,
   // results in applying the generational barrier.
   if (local_data.heap_base.in_atomic_pause()) return;
 
-  if (value_offset > 0 && age_table[value_offset] == AgeTable::Age::kOld)
+  if (value_offset > 0 && age_table.GetAge(value_offset) == AgeTable::Age::kOld)
     return;
 
   // Record slot.

test/unittests/heap/cppgc/minor-gc-unittest.cc

@@ -262,8 +262,9 @@ void InterGenerationalPointerTest(MinorGCTest* test, cppgc::Heap* heap) {
       const uintptr_t offset =
           internal_heap->caged_heap().OffsetFromAddress(young);
       // Age may be young or unknown.
-      EXPECT_NE(AgeTable::Age::kOld,
-                Heap::From(heap)->caged_heap().local_data().age_table[offset]);
+      EXPECT_NE(
+          AgeTable::Age::kOld,
+          Heap::From(heap)->caged_heap().local_data().age_table.GetAge(offset));
     }
   }