Reduce background GC allocations

src/coreclr/System.Private.CoreLib/src/System/GC.CoreCLR.cs

@@ -60,7 +60,7 @@
     public static partial class GC
     {
         [MethodImpl(MethodImplOptions.InternalCall)]
-        private static extern void GetMemoryInfo(GCMemoryInfoData data, int kind);
+        private static extern unsafe void GetMemoryInfo(GCMemoryInfoData* data, int kind);
 
         /// <summary>Gets garbage collection memory information.</summary>
         /// <returns>An object that contains information about the garbage collector's memory usage.</returns>
@@ -69,7 +69,7 @@
         /// <summary>Gets garbage collection memory information.</summary>
         /// <param name="kind">The kind of collection for which to retrieve memory information.</param>
         /// <returns>An object that contains information about the garbage collector's memory usage.</returns>
-        public static GCMemoryInfo GetGCMemoryInfo(GCKind kind)
+        public static unsafe GCMemoryInfo GetGCMemoryInfo(GCKind kind)
         {
             if ((kind < GCKind.Any) || (kind > GCKind.Background))
             {
@@ -80,8 +80,8 @@
                                           GCKind.Background));
             }
 
-            var data = new GCMemoryInfoData();
-            GetMemoryInfo(data, (int)kind);
+            GCMemoryInfoData data = default;
+            GetMemoryInfo(&data, (int)kind);
             return new GCMemoryInfo(data);
         }
 
@@ -414,7 +414,7 @@
 
 
         /// <summary>
         /// Get a count of the bytes allocated over the lifetime of the process.
         /// </summary>
         /// <param name="precise">If true, gather a precise number, otherwise gather a fairly count. Gathering a precise value triggers at a significant performance penalty.</param>
         public static long GetTotalAllocatedBytes(bool precise = false) => precise ? GetTotalAllocatedBytesPrecise() : GetTotalAllocatedBytesApproximate();
@@ -599,12 +599,12 @@
                     return true;
                 }
 
                 // We need to take a snapshot of s_notifications.Count, so that in the case that
                 // s_notifications[i].Notification() registers new notifications, we neither get rid
                 // of them nor iterate over them.
                 int count = s_notifications.Count;
 
                 // If there is no existing notifications, we won't be iterating over any and we won't
                 // be adding any new one. Also, there wasn't any added since we last invoked this
                 // method so it's safe to assume we can reset s_previousMemoryLoad.
                 if (count == 0)

src/coreclr/nativeaot/System.Private.CoreLib/src/System/GC.NativeAot.cs

@@ -764,7 +764,7 @@ public static long GetTotalAllocatedBytes(bool precise = false)
         /// <summary>Gets garbage collection memory information.</summary>
         /// <param name="kind">The kind of collection for which to retrieve memory information.</param>
         /// <returns>An object that contains information about the garbage collector's memory usage.</returns>
-        public static GCMemoryInfo GetGCMemoryInfo(GCKind kind)
+        public static unsafe GCMemoryInfo GetGCMemoryInfo(GCKind kind)
         {
             if ((kind < GCKind.Any) || (kind > GCKind.Background))
             {
@@ -775,8 +775,8 @@ public static GCMemoryInfo GetGCMemoryInfo(GCKind kind)
                                           GCKind.Background));
             }
 
-            var data = new GCMemoryInfoData();
-            RuntimeImports.RhGetMemoryInfo(ref data.GetRawData(), kind);
+            GCMemoryInfoData data = default;
+            RuntimeImports.RhGetMemoryInfo(&data, kind);
             return new GCMemoryInfo(data);
         }
 

src/coreclr/nativeaot/System.Private.CoreLib/src/System/Runtime/RuntimeImports.cs

@@ -241,7 +241,7 @@ internal struct GCHeapHardLimitInfo
 
         [MethodImpl(MethodImplOptions.InternalCall)]
         [RuntimeImport(RuntimeLibrary, "RhGetMemoryInfo")]
-        internal static extern void RhGetMemoryInfo(ref byte info, GCKind kind);
+        internal static extern unsafe void RhGetMemoryInfo(GCMemoryInfoData* data, GCKind kind);
 
         [LibraryImport(RuntimeLibrary)]
         internal static unsafe partial void RhAllocateNewArray(MethodTable* pArrayEEType, uint numElements, uint flags, void* pResult);

src/coreclr/vm/comutilnative.cpp

@@ -584,32 +584,30 @@ FCIMPL0(INT64, GCInterface::GetTotalPauseDuration)
 }
 FCIMPLEND
 
-FCIMPL2(void, GCInterface::GetMemoryInfo, Object* objUNSAFE, int kind)
+FCIMPL2(void, GCInterface::GetMemoryInfo, GCMemoryInfoData* pData, int kind)
 {
     FCALL_CONTRACT;
 
     FC_GC_POLL_NOT_NEEDED();
 
-    GCMEMORYINFODATAREF objGCMemoryInfo = (GCMEMORYINFODATAREF)(ObjectToOBJECTREF (objUNSAFE));
-
-    UINT64* genInfoRaw = (UINT64*)&(objGCMemoryInfo->generationInfo0);
-    UINT64* pauseInfoRaw = (UINT64*)&(objGCMemoryInfo->pauseDuration0);
+    UINT64* genInfoRaw = (UINT64*)&(pData->generationInfo0);
+    UINT64* pauseInfoRaw = (UINT64*)&(pData->pauseDuration0);
 
     return GCHeapUtilities::GetGCHeap()->GetMemoryInfo(
-        &(objGCMemoryInfo->highMemLoadThresholdBytes),
-        &(objGCMemoryInfo->totalAvailableMemoryBytes),
-        &(objGCMemoryInfo->lastRecordedMemLoadBytes),
-        &(objGCMemoryInfo->lastRecordedHeapSizeBytes),
-        &(objGCMemoryInfo->lastRecordedFragmentationBytes),
-        &(objGCMemoryInfo->totalCommittedBytes),
-        &(objGCMemoryInfo->promotedBytes),
-        &(objGCMemoryInfo->pinnedObjectCount),
-        &(objGCMemoryInfo->finalizationPendingCount),
-        &(objGCMemoryInfo->index),
-        &(objGCMemoryInfo->generation),
-        &(objGCMemoryInfo->pauseTimePercent),
-        (bool*)&(objGCMemoryInfo->isCompaction),
-        (bool*)&(objGCMemoryInfo->isConcurrent),
+        &(pData->highMemLoadThresholdBytes),
+        &(pData->totalAvailableMemoryBytes),
+        &(pData->lastRecordedMemLoadBytes),
+        &(pData->lastRecordedHeapSizeBytes),
+        &(pData->lastRecordedFragmentationBytes),
+        &(pData->totalCommittedBytes),
+        &(pData->promotedBytes),
+        &(pData->pinnedObjectCount),
+        &(pData->finalizationPendingCount),
+        &(pData->index),
+        &(pData->generation),
+        &(pData->pauseTimePercent),
+        (bool*)&(pData->isCompaction),
+        (bool*)&(pData->isConcurrent),
         genInfoRaw,
         pauseInfoRaw,
         kind);

src/coreclr/vm/comutilnative.h

@@ -101,7 +101,7 @@ struct GCGenerationInfo
 };
 
 #include "pshpack4.h"
-class GCMemoryInfoData : public Object
+struct GCMemoryInfoData
 {
 public:
     UINT64 highMemLoadThresholdBytes;
@@ -131,14 +131,6 @@ class GCMemoryInfoData : public Object
 };
 #include "poppack.h"
 
-#ifdef USE_CHECKED_OBJECTREFS
-typedef REF<GCMemoryInfoData> GCMEMORYINFODATA;
-typedef REF<GCMemoryInfoData> GCMEMORYINFODATAREF;
-#else // USE_CHECKED_OBJECTREFS
-typedef GCMemoryInfoData * GCMEMORYINFODATA;
-typedef GCMemoryInfoData * GCMEMORYINFODATAREF;
-#endif // USE_CHECKED_OBJECTREFS
-
 using EnumerateConfigurationValuesCallback = void (*)(void* context, void* name, void* publicKey, GCConfigurationType type, int64_t data);
 
 struct GCHeapHardLimitInfo
@@ -166,7 +158,7 @@ class GCInterface {
     static FORCEINLINE UINT64 InterlockedSub(UINT64 *pMinuend, UINT64 subtrahend);
 
     static FCDECL0(INT64,   GetTotalPauseDuration);
-    static FCDECL2(void,    GetMemoryInfo, Object* objUNSAFE, int kind);
+    static FCDECL2(void,    GetMemoryInfo, GCMemoryInfoData* pData, int kind);
     static FCDECL0(UINT32,  GetMemoryLoad);
     static FCDECL0(int,     GetGcLatencyMode);
     static FCDECL1(int,     SetGcLatencyMode, int newLatencyMode);

src/libraries/System.Private.CoreLib/src/System/Buffers/SharedArrayPool.cs

@@ -209,14 +209,14 @@ public bool Trim()
             {
                 if (!log.IsEnabled())
                 {
-                    foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets)
+                    foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets.EnumerateOnStack())
                     {
                         Array.Clear(tlsBuckets.Key);
                     }
                 }
                 else
                 {
-                    foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets)
+                    foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets.EnumerateOnStack())
                     {
                         SharedArrayPoolThreadLocalArray[] buckets = tlsBuckets.Key;
                         for (int i = 0; i < buckets.Length; i++)
@@ -241,7 +241,7 @@ public bool Trim()
                     _ => 30_000,
                 };
 
-                foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets)
+                foreach (KeyValuePair<SharedArrayPoolThreadLocalArray[], object?> tlsBuckets in _allTlsBuckets.EnumerateOnStack())
                 {
                     SharedArrayPoolThreadLocalArray[] buckets = tlsBuckets.Key;
                     for (int i = 0; i < buckets.Length; i++)

src/libraries/System.Private.CoreLib/src/System/GCMemoryInfo.cs

@@ -54,7 +54,7 @@ public enum GCKind
     };
 
     [StructLayout(LayoutKind.Sequential)]
-    internal sealed class GCMemoryInfoData
+    internal struct GCMemoryInfoData
     {
         internal long _highMemoryLoadThresholdBytes;
         internal long _totalAvailableMemoryBytes;

src/libraries/System.Private.CoreLib/src/System/Runtime/CompilerServices/ConditionalWeakTable.cs

@@ -369,8 +369,8 @@ IEnumerator<KeyValuePair<TKey, TValue>> IEnumerable<KeyValuePair<TKey, TValue>>.
 
         IEnumerator IEnumerable.GetEnumerator() => ((IEnumerable<KeyValuePair<TKey, TValue>>)this).GetEnumerator();
 
-        /// <summary>Provides an enumerator for the table.</summary>
-        private sealed class Enumerator : IEnumerator<KeyValuePair<TKey, TValue>>
+        /// <summary>The core implementation for the enumerators for the table.</summary>
+        private struct EnumeratorCore : IDisposable
         {
             // The enumerator would ideally hold a reference to the Container and the end index within that
             // container.  However, the safety of the CWT depends on the only reference to the Container being
@@ -393,7 +393,7 @@ private sealed class Enumerator : IEnumerator<KeyValuePair<TKey, TValue>>
             private int _currentIndex;                          // the current index into the container
             private KeyValuePair<TKey, TValue> _current;        // the current entry set by MoveNext and returned from Current
 
-            public Enumerator(ConditionalWeakTable<TKey, TValue> table)
+            public EnumeratorCore(ConditionalWeakTable<TKey, TValue> table)
             {
                 Debug.Assert(table != null, "Must provide a valid table");
                 Debug.Assert(Monitor.IsEntered(table._lock), "Must hold the _lock lock to construct the enumerator");
@@ -410,11 +410,6 @@ public Enumerator(ConditionalWeakTable<TKey, TValue> table)
                 _currentIndex = -1;
             }
 
-            ~Enumerator()
-            {
-                Dispose();
-            }
-
             public void Dispose()
             {
                 // Use an interlocked operation to ensure that only one thread can get access to
@@ -485,6 +480,56 @@ public KeyValuePair<TKey, TValue> Current
                     return _current;
                 }
             }
+        }
+
+        /// <summary>Provides an enumerator for the table.</summary>
+        private sealed class Enumerator : IEnumerator<KeyValuePair<TKey, TValue>>
+        {
+            private EnumeratorCore _enumeratorCore;
+
+            public Enumerator(ConditionalWeakTable<TKey, TValue> table) => _enumeratorCore = new(table);
+
+            ~Enumerator()
+            {
+                Dispose();
+            }
+
+            public void Dispose() => _enumeratorCore.Dispose();
+
+            public bool MoveNext() => _enumeratorCore.MoveNext();
+
+            public KeyValuePair<TKey, TValue> Current => _enumeratorCore.Current;
+
+            object? IEnumerator.Current => Current;
+
+            public void Reset() { }
+        }
+
+        /// <summary>Gets an enumerable used to enumerate the values on the stack without allocation.</summary>
+        // This API is unsafe to use if the user neglects to call Dispose, which is why
+        // we can only use this internally where we can enforce proper foreach usage.
+        // It's also unsafe to use in async functions, even with proper foreach, because there's
+        // a chance that the async will never continue, in which case we need to rely on the finalizer in the regular enumerator.
+        internal RefEnumerable EnumerateOnStack() => new RefEnumerable(this);
+
+        /// <summary>Provides an enumerable used to enumerate the values on the stack without allocation.</summary>
+        internal readonly ref struct RefEnumerable(ConditionalWeakTable<TKey, TValue> table)
+        {
+            public RefEnumerator GetEnumerator() => new(table);
+        }
+
+        /// <summary>Provides an enumerator for the table that can be used on the stack without allocation.</summary>
+        internal ref struct RefEnumerator : IEnumerator<KeyValuePair<TKey, TValue>>
+        {
+            private EnumeratorCore _enumeratorCore;
+
+            public RefEnumerator(ConditionalWeakTable<TKey, TValue> table) => _enumeratorCore = new(table);
+
+            public void Dispose() => _enumeratorCore.Dispose();
+
+            public bool MoveNext() => _enumeratorCore.MoveNext();
+
+            public KeyValuePair<TKey, TValue> Current => _enumeratorCore.Current;
 
             object? IEnumerator.Current => Current;
 

src/mono/System.Private.CoreLib/src/System/GC.Mono.cs

@@ -248,7 +248,7 @@ private static extern void _GetGCMemoryInfo(out long highMemoryLoadThresholdByte
 
         public static GCMemoryInfo GetGCMemoryInfo()
         {
-            var data = new GCMemoryInfoData();
+            GCMemoryInfoData data = default;
 
             _GetGCMemoryInfo(out data._highMemoryLoadThresholdBytes,
                              out data._memoryLoadBytes,