Enable TYP_STRUCT LCL_VAR/LCL_FLD call args on Windows x64 (#70777)

* Fix forward sub

* Enable folding in local morph

* Morph: TYP_STRUCT LCL_FLD

Author: SingleAccretion

src/coreclr/jit/forwardsub.cpp

@@ -659,6 +659,15 @@ bool Compiler::fgForwardSubStatement(Statement* stmt)
 
     // Quirks:
     //
+    // Don't substitute nodes "AddFinalArgsAndDetermineABIInfo" doesn't handle into struct args.
+    //
+    if (fsv.IsCallArg() && fsv.GetNode()->TypeIs(TYP_STRUCT) &&
+        !fwdSubNode->OperIs(GT_OBJ, GT_LCL_VAR, GT_LCL_FLD, GT_MKREFANY))
+    {
+        JITDUMP(" use is a struct arg; fwd sub node is not OBJ/LCL_VAR/LCL_FLD/MKREFANY\n");
+        return false;
+    }
+
     // We may sometimes lose or change a type handle. Avoid substituting if so.
     //
     // However, we allow free substitution of hardware SIMD types.

src/coreclr/jit/gentree.cpp

@@ -597,6 +597,9 @@ ClassLayout* GenTree::GetLayout(Compiler* compiler) const
         case GT_BLK:
             return AsBlk()->GetLayout();
 
+        case GT_MKREFANY:
+            return compiler->typGetObjLayout(compiler->impGetRefAnyClass());
+
         default:
             unreached();
     }

src/coreclr/jit/lclmorph.cpp

@@ -1049,10 +1049,9 @@ class LocalAddressVisitor final : public GenTreeVisitor<LocalAddressVisitor>
             return IndirTransform::None;
         }
 
-        if ((user == nullptr) || !user->OperIs(GT_ASG, GT_RETURN))
+        if ((user == nullptr) || !user->OperIs(GT_ASG, GT_CALL, GT_RETURN))
         {
-            // TODO-ADDR: call args require extra work because currently they must
-            // be wrapped in OBJ nodes so we can't replace those with local nodes.
+            // TODO-ADDR: remove unused indirections.
             return IndirTransform::None;
         }
 
@@ -1075,7 +1074,6 @@ class LocalAddressVisitor final : public GenTreeVisitor<LocalAddressVisitor>
         //
         enum class StructMatch
         {
-            Exact,
             Compatible,
             Partial
         };
@@ -1084,44 +1082,44 @@ class LocalAddressVisitor final : public GenTreeVisitor<LocalAddressVisitor>
         assert(varDsc->GetLayout() != nullptr);
 
         StructMatch match = StructMatch::Partial;
-        if (val.Offset() == 0)
+        if ((val.Offset() == 0) && ClassLayout::AreCompatible(indirLayout, varDsc->GetLayout()))
         {
-            if (indirLayout->GetClassHandle() == varDsc->GetStructHnd())
-            {
-                match = StructMatch::Exact;
-            }
-            else if (ClassLayout::AreCompatible(indirLayout, varDsc->GetLayout()))
-            {
-                match = StructMatch::Compatible;
-            }
+            match = StructMatch::Compatible;
         }
 
         // Current matrix of matches/users/types:
         //
-        // |------------|------|-------------|---------|
-        // | STRUCT     | CALL | ASG         | RETURN  |
-        // |------------|------|-------------|---------|
-        // | Exact      | None | LCL_VAR     | LCL_VAR |
-        // | Compatible | None | LCL_VAR     | LCL_VAR |
-        // | Partial    | None | OBJ/LCL_FLD | LCL_FLD |
-        // |------------|------|-------------|---------|
+        // |------------|---------|-------------|---------|
+        // | STRUCT     | CALL(*) | ASG         | RETURN  |
+        // |------------|---------|-------------|---------|
+        // | Compatible | LCL_VAR | LCL_VAR     | LCL_VAR |
+        // | Partial    | LCL_FLD | OBJ/LCL_FLD | LCL_FLD |
+        // |------------|---------|-------------|---------|
+        //
+        // * - On Windows x64 only.
         //
         // |------------|------|------|--------|----------|
         // | SIMD       | CALL | ASG  | RETURN | HWI/SIMD |
         // |------------|------|------|--------|----------|
-        // | Exact      | None | None | None   | None     |
         // | Compatible | None | None | None   | None     |
         // | Partial    | None | None | None   | None     |
         // |------------|------|------|--------|----------|
         //
         // TODO-ADDR: delete all the "None" entries and always
         // transform local nodes into LCL_VAR or LCL_FLD.
 
-        assert(indir->TypeIs(TYP_STRUCT) && user->OperIs(GT_ASG, GT_RETURN));
+        assert(indir->TypeIs(TYP_STRUCT) && user->OperIs(GT_ASG, GT_CALL, GT_RETURN));
 
         *pStructLayout = indirLayout;
 
-        if ((match == StructMatch::Exact) || (match == StructMatch::Compatible))
+        if (user->IsCall())
+        {
+#ifndef WINDOWS_AMD64_ABI
+            return IndirTransform::None;
+#endif // !WINDOWS_AMD64_ABI
+        }
+
+        if (match == StructMatch::Compatible)
         {
             return IndirTransform::LclVar;
         }

src/coreclr/jit/morph.cpp

@@ -2347,30 +2347,9 @@ void CallArgs::AddFinalArgsAndDetermineABIInfo(Compiler* comp, GenTreeCall* call
             GenTree* actualArg = argx->gtEffectiveVal(true /* Commas only */);
 
             // Here we look at "actualArg" to avoid calling "getClassSize".
-            if (actualArg->TypeGet() == TYP_STRUCT)
-            {
-                switch (actualArg->OperGet())
-                {
-                    case GT_OBJ:
-                        structSize = actualArg->AsObj()->Size();
-                        break;
-                    case GT_LCL_VAR:
-                        structSize = comp->lvaGetDesc(actualArg->AsLclVarCommon())->lvExactSize;
-                        break;
-                    case GT_MKREFANY:
-                        structSize = comp->info.compCompHnd->getClassSize(argSigClass);
-                        break;
-                    default:
-                        BADCODE("illegal argument tree: cannot determine size for ABI handling");
-                        break;
-                }
-            }
-            else
-            {
-                structSize = genTypeSize(actualArg);
-            }
+            structSize = actualArg->TypeIs(TYP_STRUCT) ? actualArg->GetLayout(comp)->GetSize() : genTypeSize(actualArg);
 
-            assert(structSize = comp->info.compCompHnd->getClassSize(argSigClass));
+            assert(structSize == comp->info.compCompHnd->getClassSize(argSigClass));
         }
 #if defined(TARGET_AMD64)
 #ifdef UNIX_AMD64_ABI
@@ -3190,15 +3169,8 @@ GenTreeCall* Compiler::fgMorphArgs(GenTreeCall* call)
             unsigned originalSize;
             if (argObj->TypeGet() == TYP_STRUCT)
             {
-                if (argObj->OperIs(GT_OBJ))
-                {
-                    originalSize = argObj->AsObj()->Size();
-                }
-                else
-                {
-                    // Must be LCL_VAR: we have a BADCODE assert for this in AddFinalArgsAndDetermineABIInfo.
-                    originalSize = lvaGetDesc(argObj->AsLclVar())->lvExactSize;
-                }
+                assert(argObj->OperIs(GT_OBJ, GT_LCL_VAR, GT_LCL_FLD));
+                originalSize = argObj->GetLayout(this)->GetSize();
             }
             else
             {