[clr-interp] Intrinic handling for Delegate Invoke methods (#118406)

* [Interpreter] Intrinic handling for Delegate Invoke methods
- Add a new opcode
- Add a new invocation method. Note that this needs to have the same copy the CallStubHeader logic as calli, as the target is not consistent from use to use

Co-authored-by: Copilot <[email protected]>
Co-authored-by: Jan Kotas <[email protected]>

Author: 3 people

src/coreclr/interpreter/compiler.cpp

@@ -2712,6 +2712,7 @@ void InterpCompiler::EmitCall(CORINFO_RESOLVED_TOKEN* pConstrainedToken, bool re
 {
     uint32_t token = getU4LittleEndian(m_ip + 1);
     bool isVirtual = (*m_ip == CEE_CALLVIRT);
+    bool isDelegateInvoke = false;
 
     CORINFO_RESOLVED_TOKEN resolvedCallToken;
     CORINFO_CALL_INFO callInfo;
@@ -2763,6 +2764,11 @@ void InterpCompiler::EmitCall(CORINFO_RESOLVED_TOKEN* pConstrainedToken, bool re
             }
         }
 
+        if (callInfo.methodFlags & CORINFO_FLG_DELEGATE_INVOKE)
+        {
+            isDelegateInvoke = true;
+        }
+
         if (callInfo.thisTransform != CORINFO_NO_THIS_TRANSFORM)
         {
             assert(pConstrainedToken != NULL);
@@ -3036,7 +3042,15 @@ void InterpCompiler::EmitCall(CORINFO_RESOLVED_TOKEN* pConstrainedToken, bool re
                     // before the call.
                     // TODO: Add null checking behavior somewhere here!
                 }
-                AddIns((isPInvoke && !isMarshaledPInvoke) ? INTOP_CALL_PINVOKE : INTOP_CALL);
+                if (isDelegateInvoke)
+                {
+                    assert(!isPInvoke && !isMarshaledPInvoke);
+                    AddIns(INTOP_CALLDELEGATE);
+                }
+                else
+                {
+                    AddIns((isPInvoke && !isMarshaledPInvoke) ? INTOP_CALL_PINVOKE : INTOP_CALL);
+                }
                 m_pLastNewIns->data[0] = GetMethodDataItemIndex(callInfo.hMethod);
                 if (isPInvoke && !isMarshaledPInvoke)
                 {

src/coreclr/interpreter/intops.def

@@ -356,6 +356,7 @@ OPDEF(INTOP_LDFLDA, "ldflda", 4, 1, 1, InterpOpInt)
 
 // Calls
 OPDEF(INTOP_CALL, "call", 4, 1, 1, InterpOpMethodHandle)
+OPDEF(INTOP_CALLDELEGATE, "call.delegate", 4, 1, 1, InterpOpMethodHandle)
 OPDEF(INTOP_CALLI, "calli", 5, 1, 2, InterpOpLdPtr)
 OPDEF(INTOP_CALLVIRT, "callvirt", 4, 1, 1, InterpOpMethodHandle)
 OPDEF(INTOP_CALL_PINVOKE, "call.pinvoke", 6, 1, 1, InterpOpMethodHandle) // inlined (no marshaling wrapper) pinvokes only

src/coreclr/vm/interpexec.cpp

@@ -14,9 +14,41 @@
 #ifdef TARGET_WASM
 void InvokeCalliStub(PCODE ftn, CallStubHeader *stubHeaderTemplate, int8_t *pArgs, int8_t *pRet);
 void InvokeCompiledMethod(MethodDesc *pMD, int8_t *pArgs, int8_t *pRet, PCODE target);
+void InvokeDelegateInvokeMethod(MethodDesc *pMDDelegateInvoke, int8_t *pArgs, int8_t *pRet, PCODE target);
 #else
 #include "callstubgenerator.h"
 
+CallStubHeader *UpdateCallStubForMethod(MethodDesc *pMD)
+{
+    CONTRACTL
+    {
+        THROWS;
+        MODE_ANY;
+        PRECONDITION(CheckPointer(pMD));
+    }
+    CONTRACTL_END
+
+    GCX_PREEMP();
+
+    CallStubGenerator callStubGenerator;
+
+    AllocMemTracker amTracker;
+    CallStubHeader *header = callStubGenerator.GenerateCallStub(pMD, &amTracker, true /* interpreterToNative */);
+
+    if (pMD->SetCallStub(header))
+    {
+        amTracker.SuppressRelease();
+    }
+    else
+    {
+        // We have lost the race for generating the header, use the one that was generated by another thread
+        // and let the amTracker release the memory of the one we generated.
+        header = pMD->GetCallStub();
+    }
+
+    return header;
+}
+
 void InvokeCompiledMethod(MethodDesc *pMD, int8_t *pArgs, int8_t *pRet, PCODE target)
 {
     CONTRACTL
@@ -32,22 +64,7 @@ void InvokeCompiledMethod(MethodDesc *pMD, int8_t *pArgs, int8_t *pRet, PCODE ta
     CallStubHeader *pHeader = pMD->GetCallStub();
     if (pHeader == NULL)
     {
-        CallStubGenerator callStubGenerator;
-        GCX_PREEMP();
-
-        AllocMemTracker amTracker;
-        pHeader = callStubGenerator.GenerateCallStub(pMD, &amTracker, true /* interpreterToNative */);
-
-        if (pMD->SetCallStub(pHeader))
-        {
-            amTracker.SuppressRelease();
-        }
-        else
-        {
-            // We have lost the race for generating the header, use the one that was generated by another thread
-            // and let the amTracker release the memory of the one we generated.
-            pHeader = pMD->GetCallStub();
-        }
+        pHeader = UpdateCallStubForMethod(pMD);
     }
 
     // Interpreter-FIXME: Potential race condition if a single CallStubHeader is reused for multiple targets.
@@ -56,6 +73,34 @@ void InvokeCompiledMethod(MethodDesc *pMD, int8_t *pArgs, int8_t *pRet, PCODE ta
     pHeader->Invoke(pHeader->Routines, pArgs, pRet, pHeader->TotalStackSize);
 }
 
+void InvokeDelegateInvokeMethod(MethodDesc *pMDDelegateInvoke, int8_t *pArgs, int8_t *pRet, PCODE target)
+{
+    CONTRACTL
+    {
+        THROWS;
+        MODE_ANY;
+        PRECONDITION(CheckPointer(pMDDelegateInvoke));
+        PRECONDITION(CheckPointer(pArgs));
+        PRECONDITION(CheckPointer(pRet));
+    }
+    CONTRACTL_END
+
+    CallStubHeader *stubHeaderTemplate = pMDDelegateInvoke->GetCallStub();
+    if (stubHeaderTemplate == NULL)
+    {
+        stubHeaderTemplate = UpdateCallStubForMethod(pMDDelegateInvoke);
+    }
+
+    // CallStubHeaders encode their destination addresses in the Routines array, so they need to be
+    // copied to a local buffer before we can actually set their target address.
+    size_t templateSize = stubHeaderTemplate->GetSize();
+    uint8_t* actualCallStub = (uint8_t*)alloca(templateSize);
+    memcpy(actualCallStub, stubHeaderTemplate, templateSize);
+    CallStubHeader *pHeader = (CallStubHeader*)actualCallStub;
+    pHeader->SetTarget(target); // The method to call
+    pHeader->Invoke(pHeader->Routines, pArgs, pRet, pHeader->TotalStackSize);
+}
+
 void InvokeCalliStub(PCODE ftn, CallStubHeader *stubHeaderTemplate, int8_t *pArgs, int8_t *pRet)
 {
     CONTRACTL
@@ -84,7 +129,6 @@ CallStubHeader *CreateNativeToInterpreterCallStub(InterpMethod* pInterpMethod)
     CallStubHeader *pHeader = VolatileLoadWithoutBarrier(&pInterpMethod->pCallStub);
     GCX_PREEMP();
 
-    AllocMemTracker amTracker;
     if (pHeader == NULL)
     {
         // Ensure that there is an interpreter thread context instance and thus an interpreter stack
@@ -1881,6 +1925,30 @@ void InterpExecMethod(InterpreterFrame *pInterpreterFrame, InterpMethodContextFr
                     break;
                 }
 
+                case INTOP_CALLDELEGATE:
+                {
+                    returnOffset = ip[1];
+                    callArgsOffset = ip[2];
+                    methodSlot = ip[3];
+
+                    // targetMethod holds a pointer to the Invoke method of the delegate, not the final actual target.
+                    targetMethod = (MethodDesc*)pMethod->pDataItems[methodSlot];
+
+                    ip += 4;
+
+                    DELEGATEREF delegateObj = LOCAL_VAR(callArgsOffset, DELEGATEREF);
+                    NULL_CHECK(delegateObj);
+                    PCODE targetAddress = delegateObj->GetMethodPtr();
+                    OBJECTREF targetMethodObj = delegateObj->GetTarget();
+                    LOCAL_VAR(callArgsOffset, OBJECTREF) = targetMethodObj;
+
+                    // TODO! Once we are investigating performance here, we may want to optimize this so that
+                    // delegate calls to interpeted methods don't have to go through the native invoke here, but for
+                    // now this should work well.
+                    InvokeDelegateInvokeMethod(targetMethod, stack + callArgsOffset, stack + returnOffset, targetAddress);
+                    break;
+                }
+
                 case INTOP_CALL:
                 {
                     returnOffset = ip[1];

src/coreclr/vm/wasm/helpers.cpp

@@ -487,3 +487,8 @@ void InvokeCompiledMethod(MethodDesc *pMD, int8_t *pArgs, int8_t *pRet, PCODE ta
 {
     PORTABILITY_ASSERT("Attempted to execute non-interpreter code from interpreter on wasm, this is not yet implemented");
 }
+
+void InvokeDelegateInvokeMethod(MethodDesc *pMDDelegateInvoke, int8_t *pArgs, int8_t *pRet, PCODE target)
+{
+    PORTABILITY_ASSERT("Attempted to execute non-interpreter code from interpreter on wasm, this is not yet implemented");
+}