[interp] Reverse p/invoke (#119133)

* Implements GetStubContext intrinsic in the interpreter
* For interpreted reverse p/invoke stubs, we ensure that TheUMEntryPrestubWorker runs for every invocation so that it can capture the hidden argument in a thread local. The interpreter can then recover the hidden argument from the thread local in order to offer it via GetStubContext later.
* Upon entry to interpreted methods that accept a hidden argument, we store the hidden argument (that was saved by the um entry prestub worker) into a special local that can be read from later by calls to GetStubContext.

Author: kg

src/coreclr/interpreter/compiler.cpp

@@ -1288,6 +1288,7 @@ InterpCompiler::InterpCompiler(COMP_HANDLE compHnd,
                                 CORINFO_METHOD_INFO* methodInfo)
     : m_stackmapsByClass(FreeInterpreterStackMap)
     , m_pInitLocalsIns(nullptr)
+    , m_hiddenArgumentVar(-1)
     , m_globalVarsWithRefsStackTop(0)
 {
     m_genericLookupToDataItemIndex.Init(&m_dataItems, this);
@@ -2306,6 +2307,16 @@ bool InterpCompiler::EmitNamedIntrinsicCall(NamedIntrinsic ni, CORINFO_CLASS_HAN
             AddIns(INTOP_THROW_PNSE);
             return true;
 
+        case NI_System_StubHelpers_GetStubContext:
+        {
+            assert(m_hiddenArgumentVar >= 0);
+            AddIns(INTOP_MOV_P);
+            PushStackType(StackTypeI, NULL);
+            m_pLastNewIns->SetSVar(m_hiddenArgumentVar);
+            m_pLastNewIns->SetDVar(m_pStackPointer[-1].var);
+            return true;
+        }
+
         case NI_System_Runtime_CompilerServices_StaticsHelpers_VolatileReadAsByref:
         {
             CHECK_STACK(1);
@@ -2852,7 +2863,7 @@ void InterpCompiler::EmitCall(CORINFO_RESOLVED_TOKEN* pConstrainedToken, bool re
             //  intrinsics for the recursive call. Otherwise we will just recurse infinitely and overflow stack.
             //  This expansion can produce value that is inconsistent with the value seen by JIT/R2R code that can
             //  cause user code to misbehave. This is by design. One-off method Interpretation is for internal use only.
-            bool isMustExpand = (callInfo.hMethod == m_methodHnd);
+            bool isMustExpand = (callInfo.hMethod == m_methodHnd) || (ni == NI_System_StubHelpers_GetStubContext);
             if ((InterpConfig.InterpMode() == 3) || isMustExpand)
             {
                 if (EmitNamedIntrinsicCall(ni, resolvedCallToken.hClass, callInfo.hMethod, callInfo.sig))
@@ -3697,6 +3708,17 @@ void InterpCompiler::GenerateCode(CORINFO_METHOD_INFO* methodInfo)
     // adding an opcode.
     AddIns(INTOP_SAFEPOINT);
 
+    CORJIT_FLAGS corJitFlags;
+    DWORD jitFlagsSize = m_compHnd->getJitFlags(&corJitFlags, sizeof(corJitFlags));
+    assert(jitFlagsSize == sizeof(corJitFlags));
+
+    if (corJitFlags.IsSet(CORJIT_FLAGS::CORJIT_FLAG_PUBLISH_SECRET_PARAM))
+    {
+        m_hiddenArgumentVar = CreateVarExplicit(InterpTypeI, NULL, sizeof(void *));
+        AddIns(INTOP_STORESTUBCONTEXT);
+        m_pLastNewIns->SetDVar(m_hiddenArgumentVar);
+    }
+
     CorInfoInitClassResult initOnFunctionStart = m_compHnd->initClass(NULL, NULL, METHOD_BEING_COMPILED_CONTEXT());
     if ((initOnFunctionStart & CORINFO_INITCLASS_USE_HELPER) == CORINFO_INITCLASS_USE_HELPER)
     {

src/coreclr/interpreter/compiler.h

@@ -514,6 +514,10 @@ class InterpCompiler
     int32_t m_currentILOffset;
     InterpInst* m_pInitLocalsIns;
 
+    // If the method has a hidden argument, GenerateCode allocates a var to store it and
+    //  populates the var at method entry
+    int32_t m_hiddenArgumentVar;
+
     // Table of mappings of leave instructions to the first finally call island the leave
     // needs to execute.
     TArray<LeavesTableEntry> m_leavesTable;

src/coreclr/interpreter/interpretershared.h

@@ -38,7 +38,10 @@ struct InterpMethod
     bool initLocals;
     bool unmanagedCallersOnly;
 
-    InterpMethod(CORINFO_METHOD_HANDLE methodHnd, int32_t argsSize, int32_t allocaSize, void** pDataItems, bool initLocals, bool unmanagedCallersOnly)
+    InterpMethod(
+        CORINFO_METHOD_HANDLE methodHnd, int32_t argsSize, int32_t allocaSize,
+        void** pDataItems, bool initLocals, bool unmanagedCallersOnly
+    )
     {
 #if DEBUG
         this->self = this;

src/coreclr/interpreter/intops.def

@@ -417,6 +417,7 @@ OPDEF(INTOP_COMPARE_EXCHANGE_I4, "compare.exchange.i4", 5, 1, 3, InterpOpNoArgs)
 OPDEF(INTOP_COMPARE_EXCHANGE_I8, "compare.exchange.i8", 5, 1, 3, InterpOpNoArgs)
 OPDEF(INTOP_EXCHANGE_I4, "exchange.i4", 4, 1, 2, InterpOpNoArgs)
 OPDEF(INTOP_EXCHANGE_I8, "exchange.i8", 4, 1, 2, InterpOpNoArgs)
+OPDEF(INTOP_STORESTUBCONTEXT, "storestubcontext", 2, 1, 0, InterpOpNoArgs)
 
 // All instructions after this point are IROPS, instructions that are not emitted/executed
 OPDEF(INTOP_NOP, "nop", 1, 0, 0, InterpOpNoArgs)

src/coreclr/interpreter/intrinsics.cpp

@@ -44,6 +44,8 @@ NamedIntrinsic GetNamedIntrinsic(COMP_HANDLE compHnd, CORINFO_METHOD_HANDLE comp
         {
             if (!strcmp(methodName, "NextCallReturnAddress"))
                 return NI_System_StubHelpers_NextCallReturnAddress;
+            else if (!strcmp(methodName, "GetStubContext"))
+                return NI_System_StubHelpers_GetStubContext;
         }
     }
     else if (!strcmp(namespaceName, "System.Numerics"))

src/coreclr/vm/dllimportcallback.cpp

@@ -210,6 +210,19 @@ VOID CallbackOnCollectedDelegate(UMEntryThunkData* pEntryThunkData)
     EEPOLICY_HANDLE_FATAL_ERROR_WITH_MESSAGE(COR_E_FAILFAST, message.GetUnicode());
 }
 
+#ifdef FEATURE_INTERPRETER
+PLATFORM_THREAD_LOCAL UMEntryThunkData * t_MostRecentUMEntryThunkData;
+
+UMEntryThunkData * GetMostRecentUMEntryThunkData()
+{
+    LIMITED_METHOD_CONTRACT;
+
+    UMEntryThunkData * result = t_MostRecentUMEntryThunkData;
+    t_MostRecentUMEntryThunkData = nullptr;
+    return result;
+}
+#endif
+
 PCODE TheUMEntryPrestubWorker(UMEntryThunkData * pUMEntryThunkData)
 {
     STATIC_CONTRACT_THROWS;
@@ -222,6 +235,15 @@ PCODE TheUMEntryPrestubWorker(UMEntryThunkData * pUMEntryThunkData)
         CREATETHREAD_IF_NULL_FAILFAST(pThread, W("Failed to setup new thread during reverse P/Invoke"));
     }
 
+#ifdef FEATURE_INTERPRETER
+    PCODE pInterpreterTarget = pUMEntryThunkData->GetInterpreterTarget();
+    if (pInterpreterTarget != (PCODE)0)
+    {
+        t_MostRecentUMEntryThunkData = pUMEntryThunkData;
+        return pInterpreterTarget;
+    }
+#endif // FEATURE_INTERPRETER
+
     // Verify the current thread isn't in COOP mode.
     if (pThread->PreemptiveGCDisabled())
         ReversePInvokeBadTransition();
@@ -263,7 +285,7 @@ UMEntryThunkData* UMEntryThunkData::CreateUMEntryThunk()
         LoaderAllocator *pLoaderAllocator = SystemDomain::GetGlobalLoaderAllocator();
         AllocMemTracker amTracker;
         AllocMemTracker *pamTracker = &amTracker;
-       
+
         pData = (UMEntryThunkData *)pamTracker->Track(pLoaderAllocator->GetLowFrequencyHeap()->AllocMem(S_SIZE_T(sizeof(UMEntryThunkData))));
         UMEntryThunk* pThunk = (UMEntryThunk*)pamTracker->Track(pLoaderAllocator->GetNewStubPrecodeHeap()->AllocStub());
 #ifdef FEATURE_PERFMAP

src/coreclr/vm/dllimportcallback.h

@@ -134,7 +134,7 @@ class UMEntryThunk : private StubPrecode
     PTR_UMEntryThunkData GetData() const
     {
         LIMITED_METHOD_CONTRACT;
-        
+
         return dac_cast<PTR_UMEntryThunkData>(GetSecretParam());
     }
 };
@@ -149,6 +149,13 @@ class UMEntryThunkData
     // but it may not have the correct value.
     PCODE                   m_pManagedTarget;
 
+#ifdef FEATURE_INTERPRETER
+    // InterpreterPrecode to tailcall if the target is interpreted. This allows TheUMEntryPrestubWorker
+    // stash the hidden argument in a thread static and avoid collision with the hidden argument
+    // used by InterpreterPrecode.
+    Volatile<PCODE>         m_pInterpretedTarget;
+#endif
+
     // This is used for debugging and profiling.
     PTR_MethodDesc          m_pMD;
 
@@ -200,6 +207,9 @@ class UMEntryThunkData
         m_pManagedTarget = pManagedTarget;
         m_pObjectHandle     = pObjectHandle;
         m_pUMThunkMarshInfo = pUMThunkMarshInfo;
+#ifdef FEATURE_INTERPRETER
+        m_pInterpretedTarget = (PCODE)0;
+#endif
 
         m_pMD = pMD;
 
@@ -214,7 +224,15 @@ class UMEntryThunkData
 
     void Terminate();
 
-    VOID RunTimeInit()
+#ifdef FEATURE_INTERPRETER
+    PCODE GetInterpreterTarget()
+    {
+        LIMITED_METHOD_CONTRACT;
+        return m_pInterpretedTarget;
+    }
+#endif
+
+    void RunTimeInit()
     {
         STANDARD_VM_CONTRACT;
 
@@ -227,7 +245,27 @@ class UMEntryThunkData
         if (m_pObjectHandle == NULL && m_pManagedTarget == (TADDR)0)
             m_pManagedTarget = m_pMD->GetMultiCallableAddrOfCode();
 
-        m_pUMEntryThunk->SetTargetUnconditional(m_pUMThunkMarshInfo->GetExecStubEntryPoint());
+        PCODE entryPoint = m_pUMThunkMarshInfo->GetExecStubEntryPoint();
+
+#ifdef FEATURE_INTERPRETER
+        // For interpreted stubs we need to ensure that TheUMEntryPrestubWorker runs for every
+        // unmanaged-to-managed invocation in order to populate the TLS variable every time.
+        auto stubKind = RangeSectionStubManager::GetStubKind(entryPoint);
+        if (stubKind == STUB_CODE_BLOCK_STUBPRECODE)
+        {
+            StubPrecode* pPrecode = Precode::GetPrecodeFromEntryPoint(entryPoint)->AsStubPrecode();
+            if (pPrecode->GetType() == PRECODE_INTERPRETER)
+            {
+                m_pInterpretedTarget = entryPoint;
+                entryPoint = (PCODE)0;
+            }
+        }
+
+        if (entryPoint != (PCODE)0)
+#endif // FEATURE_INTERPRETER
+        {
+            m_pUMEntryThunk->SetTargetUnconditional(entryPoint);
+        }
 
 #ifdef _DEBUG
         m_state = kRunTimeInited;

src/coreclr/vm/interpexec.cpp

@@ -530,6 +530,8 @@ void PrepareInitialCode(MethodDesc *pMD)
     PAL_ENDTRY
 }
 
+UMEntryThunkData * GetMostRecentUMEntryThunkData();
+
 void InterpExecMethod(InterpreterFrame *pInterpreterFrame, InterpMethodContextFrame *pFrame, InterpThreadContext *pThreadContext, ExceptionClauseArgs *pExceptionClauseArgs)
 {
     CONTRACTL
@@ -610,6 +612,10 @@ void InterpExecMethod(InterpreterFrame *pInterpreterFrame, InterpMethodContextFr
                     MemoryBarrier();
                     ip++;
                     break;
+                case INTOP_STORESTUBCONTEXT:
+                    LOCAL_VAR(ip[1], void*) = GetMostRecentUMEntryThunkData();
+                    ip += 2;
+                    break;
                 case INTOP_LDC_I4:
                     LOCAL_VAR(ip[1], int32_t) = ip[2];
                     ip += 3;