JIT: generalize escape analysis delegate invoke expansion slightly (#116099)

Expand the invoke when the delegate object address has been substituted directly.

Addresses example in #84872.

Author: AndyAyersMS

src/coreclr/jit/objectalloc.cpp

@@ -2456,25 +2456,28 @@ void ObjectAllocator::RewriteUses()
                     CallArg* const thisArg      = call->gtArgs.GetThisArg();
                     GenTree* const delegateThis = thisArg->GetNode();
 
-                    if (delegateThis->OperIs(GT_LCL_VAR))
+                    if (delegateThis->OperIs(GT_LCL_VAR, GT_LCL_ADDR))
                     {
                         GenTreeLclVarCommon* const lcl = delegateThis->AsLclVarCommon();
+                        bool const                 isStackAllocatedDelegate =
+                            delegateThis->OperIs(GT_LCL_ADDR) || m_allocator->DoesLclVarPointToStack(lcl->GetLclNum());
 
-                        if (m_allocator->DoesLclVarPointToStack(lcl->GetLclNum()))
+                        if (isStackAllocatedDelegate)
                         {
                             JITDUMP("Expanding delegate invoke [%06u]\n", m_compiler->dspTreeID(call));
+
                             // Expand the delgate invoke early, so that physical promotion has
                             // a chance to promote the delegate fields.
                             //
                             // Note the instance field may also be stack allocatable (someday)
                             //
-                            GenTree* const cloneThis      = m_compiler->gtClone(lcl);
+                            GenTree* const cloneThis      = m_compiler->gtClone(lcl, /* complexOk */ true);
                             unsigned const instanceOffset = m_compiler->eeGetEEInfo()->offsetOfDelegateInstance;
                             GenTree* const newThisAddr =
                                 m_compiler->gtNewOperNode(GT_ADD, TYP_I_IMPL, cloneThis,
                                                           m_compiler->gtNewIconNode(instanceOffset, TYP_I_IMPL));
 
-                            // For now assume the instance is heap...
+                            // For now assume the instance field is on the heap...
                             //
                             GenTree* const newThis = m_compiler->gtNewIndir(TYP_REF, newThisAddr);
                             thisArg->SetEarlyNode(newThis);

src/tests/JIT/opt/ObjectStackAllocation/Delegates.cs

@@ -34,6 +34,17 @@ static AllocationKind StackAllocation()
         return expectedAllocationKind;
     }
 
+    static AllocationKind HeapAllocation()
+    {
+        AllocationKind expectedAllocationKind = AllocationKind.Heap;
+        if (GCStressEnabled())
+        {
+            Console.WriteLine("GCStress is enabled");
+            expectedAllocationKind = AllocationKind.Undefined;
+        }
+        return expectedAllocationKind;
+    }
+
     static int CallTestAndVerifyAllocation(Test test, int expectedResult, AllocationKind expectedAllocationsKind, bool throws = false)
     {
         string methodName = test.Method.Name;
@@ -122,4 +133,21 @@ public static int Test1()
         RunTest1();
         return CallTestAndVerifyAllocation(RunTest1, 100, StackAllocation());
     }
+
+    // Here the delegate gets stack allocated, but not the closure.
+    // With PGO the delegate is also inlined.
+    //
+    [MethodImpl(MethodImplOptions.NoInlining)]
+    static int RunTest2Inner(int a) => InvokeFunc(x => x + a);
+
+    static int InvokeFunc(Func<int, int> func) => func(101);
+
+    [MethodImpl(MethodImplOptions.NoInlining)]
+    static int RunTest2() => RunTest2Inner(-1);
+
+    [Fact]
+    public static int Test2()
+    {
+        return CallTestAndVerifyAllocation(RunTest2, 100, HeapAllocation());
+    }
 }