Fix passing Vector3s from the interpreter to native code

src/coreclr/interpreter/compiler.cpp

@@ -2848,6 +2848,11 @@ void InterpCompiler::EmitCall(CORINFO_RESOLVED_TOKEN* pConstrainedToken, bool re
                 m_pLastNewIns->SetDVar(pThisStackInfo->var);
             }
         }
+
+#if DEBUG
+        if (InterpConfig.InterpHaltOnCall().contains(m_compHnd, resolvedCallToken.hMethod, resolvedCallToken.hClass, nullptr))
+            assert(!"HaltOnCall");
+#endif
     }
 
     if (newObj && (callInfo.classFlags & CORINFO_FLG_VAROBJSIZE))
@@ -5315,7 +5320,7 @@ void InterpCompiler::GenerateCode(CORINFO_METHOD_INFO* methodInfo)
                 int byrefOfTypedRefVar = m_pStackPointer[-1].var;
                 m_pLastNewIns->SetDVar(byrefOfTypedRefVar);
                 m_pStackPointer--;
-                
+
                 AddIns(GetStindForType(InterpTypeByRef));
                 m_pLastNewIns->data[0] = OFFSETOF__CORINFO_TypedReference__dataPtr;
                 m_pLastNewIns->SetSVars2(byrefOfTypedRefVar, addressVar);
@@ -5380,7 +5385,7 @@ void InterpCompiler::GenerateCode(CORINFO_METHOD_INFO* methodInfo)
                         m_pLastNewIns->data[2] = 0;
                         m_pLastNewIns->SetSVar(typedByRefVar);
                         m_pLastNewIns->SetDVar(classHandleVar);
-                        
+
                         AddIns(INTOP_CALL_HELPER_P_S);
                         m_pLastNewIns->data[0] = GetDataForHelperFtn(CORINFO_HELP_TYPEHANDLE_TO_RUNTIMETYPEHANDLE_MAYBENULL);
                         m_pLastNewIns->SetSVar(classHandleVar);

src/coreclr/interpreter/interpconfigvalues.h

@@ -20,6 +20,7 @@
 #endif
 
 RELEASE_CONFIG_METHODSET(Interpreter, "Interpreter")
+CONFIG_METHODSET(InterpHaltOnCall, "InterpHaltOnCall"); // Assert in the compiler when compiling a call to these method(s)
 CONFIG_METHODSET(InterpHalt, "InterpHalt");
 CONFIG_METHODSET(InterpDump, "InterpDump");
 CONFIG_INTEGER(InterpList, "InterpList", 0); // List the methods which are compiled by the interpreter JIT

src/coreclr/vm/callstubgenerator.cpp

@@ -1685,17 +1685,28 @@ void CallStubGenerator::ProcessArgument(ArgIterator *pArgIt, ArgLocDesc& argLocD
     // we always process single argument passed by reference using single routine.
     if (pArgIt != NULL && pArgIt->IsArgPassedByRef())
     {
+        int unalignedArgSize = pArgIt->GetArgSize();
+        // For the interpreter-to-native transition we need to make sure that we properly align the offsets
+        //  to interpreter stack slots. Otherwise a VT of i.e. size 12 will misalign the stack offset during
+        //  loads and we will start loading garbage into registers.
+        // We don't need to do this for native-to-interpreter transitions because the Store_Ref_xxx helpers
+        //  automatically do alignment of the stack offset themselves when updating the stack offset,
+        //  and if we were to pass them aligned sizes they would potentially read bytes past the end of the VT.
+        int alignedArgSize = m_interpreterToNative
+            ? ALIGN_UP(unalignedArgSize, 8)
+            : unalignedArgSize;
+
         if (argLocDesc.m_cGenReg == 1)
         {
             pRoutines[m_routineIndex++] = GetGPRegRefRoutine(argLocDesc.m_idxGenReg);
-            pRoutines[m_routineIndex++] = pArgIt->GetArgSize();
+            pRoutines[m_routineIndex++] = alignedArgSize;
             m_r1 = NoRange;
         }
         else
         {
             _ASSERTE(argLocDesc.m_byteStackIndex != -1);
             pRoutines[m_routineIndex++] = GetStackRefRoutine();
-            pRoutines[m_routineIndex++] = ((int64_t)pArgIt->GetArgSize() << 32) | argLocDesc.m_byteStackIndex;
+            pRoutines[m_routineIndex++] = ((int64_t)alignedArgSize << 32) | argLocDesc.m_byteStackIndex;
             m_s1 = NoRange;
         }
     }