Remove AggressiveOpt from CastHelpers.LdelemaRef/StelemRef

src/coreclr/System.Private.CoreLib/src/System/Runtime/CompilerServices/CastHelpers.cs

@@ -411,7 +411,6 @@ internal struct ArrayElement
         [DebuggerHidden]
         [StackTraceHidden]
         [DebuggerStepThrough]
-        [MethodImpl(MethodImplOptions.AggressiveOptimization)]
         private static ref object? LdelemaRef(Array array, nint index, void* type)
         {
             // this will throw appropriate exceptions if array is null or access is out of range.
@@ -427,7 +426,6 @@ internal struct ArrayElement
         [DebuggerHidden]
         [StackTraceHidden]
         [DebuggerStepThrough]
-        [MethodImpl(MethodImplOptions.AggressiveOptimization)]
         private static void StelemRef(Array array, nint index, object? obj)
         {
             // this will throw appropriate exceptions if array is null or access is out of range.

src/coreclr/vm/appdomain.cpp

@@ -1380,6 +1380,8 @@ void SystemDomain::LoadBaseSystemClasses()
         g_pWeakReferenceClass = CoreLibBinder::GetClass(CLASS__WEAKREFERENCE);
         g_pWeakReferenceOfTClass = CoreLibBinder::GetClass(CLASS__WEAKREFERENCEGENERIC);
 
+        g_pCastHelpers = CoreLibBinder::GetClass(CLASS__CASTHELPERS);
+
     #ifdef FEATURE_COMINTEROP
         if (g_pConfig->IsBuiltInCOMSupported())
         {

src/coreclr/vm/callcounting.cpp

@@ -660,6 +660,13 @@ bool CallCountingManager::SetCodeEntryPoint(
             CallCount callCountThreshold = g_pConfig->TieredCompilation_CallCountThreshold();
             _ASSERTE(callCountThreshold != 0);
 
+            // Let's tier up all cast helpers faster than other methods. This is because we want to import them as
+            // direct calls in codegen and they need to be promoted earlier than their callers.
+            if (methodDesc->GetMethodTable() == g_pCastHelpers)
+            {
+                callCountThreshold = max(1, (CallCount)(callCountThreshold / 2));
+            }
+
             NewHolder<CallCountingInfo> callCountingInfoHolder = new CallCountingInfo(activeCodeVersion, callCountThreshold);
             callCountingInfoByCodeVersionHash.Add(callCountingInfoHolder);
             callCountingInfo = callCountingInfoHolder.Extract();

src/coreclr/vm/ecall.cpp

@@ -137,27 +137,12 @@ void ECall::PopulateManagedCastHelpers()
     pDest = pMD->GetMultiCallableAddrOfCode();
     SetJitHelperFunction(CORINFO_HELP_UNBOX, pDest);
 
-    // Array element accessors are more perf sensitive than other managed helpers and indirection
-    // costs introduced by PreStub could be noticeable (7% to 30% depending on platform).
-    // Other helpers are either more complex, less common, or have their trivial case inlined by the JIT,
-    // so indirection is not as big concern.
-    // We JIT-compile the following helpers eagerly here to avoid indirection costs.
-
-    //TODO: revise if this specialcasing is still needed when crossgen supports tailcall optimizations
-    //      see: https://github.com/dotnet/runtime/issues/5857
-
     pMD = CoreLibBinder::GetMethod((BinderMethodID)(METHOD__CASTHELPERS__STELEMREF));
-    pMD->DoPrestub(NULL);
-    // This helper is marked AggressiveOptimization and its native code is in its final form.
-    // Get the code directly to avoid PreStub indirection.
-    pDest = pMD->GetNativeCode();
+    pDest = pMD->GetMultiCallableAddrOfCode();
     SetJitHelperFunction(CORINFO_HELP_ARRADDR_ST, pDest);
 
     pMD = CoreLibBinder::GetMethod((BinderMethodID)(METHOD__CASTHELPERS__LDELEMAREF));
-    pMD->DoPrestub(NULL);
-    // This helper is marked AggressiveOptimization and its native code is in its final form.
-    // Get the code directly to avoid PreStub indirection.
-    pDest = pMD->GetNativeCode();
+    pDest = pMD->GetMultiCallableAddrOfCode();
     SetJitHelperFunction(CORINFO_HELP_LDELEMA_REF, pDest);
 }
 

src/coreclr/vm/jitinterface.cpp

@@ -10614,6 +10614,14 @@ void* CEEJitInfo::getHelperFtn(CorInfoHelpFunc    ftnNum,         /* IN  */
         }
 #endif
 
+        // Check if we already have a cached address of the final target
+        static LPVOID hlpFinalTierAddrTable[DYNAMIC_CORINFO_HELP_COUNT] = {};
+        LPVOID finalTierAddr = hlpFinalTierAddrTable[dynamicFtnNum];
+        if (finalTierAddr != NULL)
+        {
+            return finalTierAddr;
+        }
+
         if (dynamicFtnNum == DYNAMIC_CORINFO_HELP_ISINSTANCEOFINTERFACE ||
             dynamicFtnNum == DYNAMIC_CORINFO_HELP_ISINSTANCEOFANY ||
             dynamicFtnNum == DYNAMIC_CORINFO_HELP_ISINSTANCEOFARRAY ||
@@ -10623,10 +10631,41 @@ void* CEEJitInfo::getHelperFtn(CorInfoHelpFunc    ftnNum,         /* IN  */
             dynamicFtnNum == DYNAMIC_CORINFO_HELP_CHKCASTINTERFACE ||
             dynamicFtnNum == DYNAMIC_CORINFO_HELP_CHKCASTCLASS ||
             dynamicFtnNum == DYNAMIC_CORINFO_HELP_CHKCASTCLASS_SPECIAL ||
-            dynamicFtnNum == DYNAMIC_CORINFO_HELP_UNBOX)
+            dynamicFtnNum == DYNAMIC_CORINFO_HELP_UNBOX ||
+            dynamicFtnNum == DYNAMIC_CORINFO_HELP_ARRADDR_ST ||
+            dynamicFtnNum == DYNAMIC_CORINFO_HELP_LDELEMA_REF)
         {
             Precode* pPrecode = Precode::GetPrecodeFromEntryPoint((PCODE)hlpDynamicFuncTable[dynamicFtnNum].pfnHelper);
             _ASSERTE(pPrecode->GetType() == PRECODE_FIXUP);
+
+            MethodDesc* helperMD = pPrecode->GetMethodDesc();
+            _ASSERT(helperMD != nullptr);
+
+            // Check if the target MethodDesc is already jitted to its final Tier
+            // so we no longer need to use indirections and can emit a direct call instead.
+            NativeCodeVersion::OptimizationTier tier;
+
+            // Avoid taking the lock for foreground jit compilations
+            if (TieredCompilationManager::IsBackgroundWorkerActive())
+            {
+                {
+                    CodeVersionManager::LockHolder codeVersioningLockHolder;
+                    NativeCodeVersion activeCodeVersion = helperMD->GetCodeVersionManager()->GetActiveILCodeVersion(helperMD)
+                        .GetActiveNativeCodeVersion(helperMD);
+                    tier = activeCodeVersion.GetOptimizationTier();
+                }
+
+                if (tier == NativeCodeVersion::OptimizationTier::OptimizationTier1 ||
+                    tier == NativeCodeVersion::OptimizationTier::OptimizationTierOptimized)
+                {
+                    finalTierAddr = (LPVOID)GetEEFuncEntryPoint(hlpDynamicFuncTable[dynamicFtnNum].pfnHelper);
+                    _ASSERT(finalTierAddr != NULL);
+                    // Cache it for future uses to avoid taking the lock again.
+                    hlpFinalTierAddrTable[dynamicFtnNum] = finalTierAddr;
+                    return finalTierAddr;
+                }
+            }
+
             *ppIndirection = ((FixupPrecode*)pPrecode)->GetTargetSlot();
             return NULL;
         }

src/coreclr/vm/tieredcompilation.cpp

@@ -384,6 +384,12 @@ bool TieredCompilationManager::TryScheduleBackgroundWorkerWithoutGCTrigger_Locke
     return false; // it's the caller's responsibility to call CreateBackgroundWorker() after leaving the GC_NOTRIGGER region
 }
 
+bool TieredCompilationManager::IsBackgroundWorkerActive()
+{
+    WRAPPER_NO_CONTRACT;
+    return s_isBackgroundWorkerRunning;
+}
+
 void TieredCompilationManager::CreateBackgroundWorker()
 {
     CONTRACTL

src/coreclr/vm/tieredcompilation.h

@@ -57,6 +57,7 @@ class TieredCompilationManager
 public:
     static bool TryScheduleBackgroundWorkerWithoutGCTrigger_Locked();
     static void CreateBackgroundWorker();
+    static bool IsBackgroundWorkerActive();
 private:
     static DWORD WINAPI BackgroundWorkerBootstrapper0(LPVOID args);
     static void BackgroundWorkerBootstrapper1(LPVOID args);

src/coreclr/vm/vars.cpp

@@ -106,6 +106,8 @@ GVAL_IMPL_INIT(DWORD, g_debuggerWordTLSIndex, TLS_OUT_OF_INDEXES);
 #endif
 GVAL_IMPL_INIT(DWORD, g_TlsIndex, TLS_OUT_OF_INDEXES);
 
+MethodTable* g_pCastHelpers;
+
 #ifndef DACCESS_COMPILE
 
 // <TODO> @TODO - PROMOTE. </TODO>

src/coreclr/vm/vars.hpp

@@ -401,6 +401,8 @@ EXTERN OBJECTHANDLE         g_pPreallocatedExecutionEngineException;
 // we use this as a dummy object to indicate free space in the handle tables -- this object is never visible to the world
 EXTERN OBJECTHANDLE         g_pPreallocatedSentinelObject;
 
+EXTERN MethodTable*         g_pCastHelpers;
+
 GPTR_DECL(Thread,g_pFinalizerThread);
 GPTR_DECL(Thread,g_pSuspensionThread);