ARM/ARM64: Optimize virtual call stub for R2R and JIT

src/coreclr/src/jit/codegenarmarch.cpp

@@ -2520,13 +2520,14 @@ void CodeGen::genCallInstruction(GenTreeCall* call)
                     INDEBUG_LDISASM_COMMA(sigInfo) nullptr, // addr
                     retSize MULTIREG_HAS_SECOND_GC_RET_ONLY_ARG(secondRetSize), ilOffset, target->GetRegNum());
     }
-#if defined(FEATURE_READYTORUN_COMPILER) && defined(TARGET_ARMARCH)
-    else if (call->IsR2RRelativeIndir())
+    else if (call->IsR2ROrVirtualStubRelativeIndir())
     {
         // Generate a direct call to a non-virtual user defined or helper method
         assert(callType == CT_HELPER || callType == CT_USER_FUNC);
-        assert(call->gtEntryPoint.accessType == IAT_PVALUE);
+        assert(((call->IsR2RRelativeIndir()) && (call->gtEntryPoint.accessType == IAT_PVALUE)) ||
+               ((call->IsVirtualStubRelativeIndir()) && (call->gtEntryPoint.accessType == IAT_VALUE)));
         assert(call->gtControlExpr == nullptr);
+        assert(!call->IsTailCall());
 
         regNumber tmpReg = call->GetSingleTempReg();
         GetEmitter()->emitIns_R_R(ins_Load(TYP_I_IMPL), emitActualTypeSize(TYP_I_IMPL), tmpReg, REG_R2R_INDIRECT_PARAM);
@@ -2540,7 +2541,6 @@ void CodeGen::genCallInstruction(GenTreeCall* call)
                     INDEBUG_LDISASM_COMMA(sigInfo) nullptr, // addr
                     retSize MULTIREG_HAS_SECOND_GC_RET_ONLY_ARG(secondRetSize), ilOffset, tmpReg);
     }
-#endif // FEATURE_READYTORUN_COMPILER && TARGET_ARMARCH
     else
     {
         // Generate a direct call to a non-virtual user defined or helper method

src/coreclr/src/jit/gentree.h

@@ -4202,6 +4202,16 @@ struct GenTreeCall final : public GenTree
         return (gtFlags & GTF_CALL_INLINE_CANDIDATE) != 0;
     }
 
+    bool IsR2ROrVirtualStubRelativeIndir()
+    {
+#if defined(FEATURE_READYTORUN_COMPILER) && defined(TARGET_ARMARCH)
+        bool isVirtualStub = (gtFlags & GTF_CALL_VIRT_KIND_MASK) == GTF_CALL_VIRT_STUB;
+        return ((IsR2RRelativeIndir()) || (isVirtualStub && (IsVirtualStubRelativeIndir())));
+#else
+        return false;
+#endif // FEATURE_READYTORUN_COMPILER && TARGET_ARMARCH
+    }
+
     bool HasNonStandardAddedArgs(Compiler* compiler) const;
     int GetNonStandardAddedArgCount(Compiler* compiler) const;
 

src/coreclr/src/jit/lower.cpp

@@ -3489,6 +3489,9 @@ GenTree* Lowering::LowerDirectCall(GenTreeCall* call)
         {
             bool isR2RRelativeIndir = false;
 #if defined(FEATURE_READYTORUN_COMPILER) && defined(TARGET_ARMARCH)
+            // Skip inserting the indirection node to load the address that is already
+            // computed in REG_R2R_INDIRECT_PARAM as a hidden parameter. Instead during the
+            // codegen, just load the call target from REG_R2R_INDIRECT_PARAM.
             isR2RRelativeIndir = call->IsR2RRelativeIndir();
 #endif // FEATURE_READYTORUN_COMPILER && TARGET_ARMARCH
 
@@ -4529,7 +4532,21 @@ GenTree* Lowering::LowerVirtualStubCall(GenTreeCall* call)
         }
         else
         {
-            result = Ind(addr);
+
+            bool shouldOptimizeVirtualStubCall = false;
+#if defined(FEATURE_READYTORUN_COMPILER) && defined(TARGET_ARMARCH)
+            // Skip inserting the indirection node to load the address that is already
+            // computed in REG_R2R_INDIRECT_PARAM as a hidden parameter. Instead during the
+            // codegen, just load the call target from REG_R2R_INDIRECT_PARAM.
+            // However, for tail calls, the call target is always computed in RBM_FASTTAILCALL_TARGET
+            // and so do not optimize virtual stub calls for such cases.
+            shouldOptimizeVirtualStubCall = !call->IsTailCall();
+#endif // FEATURE_READYTORUN_COMPILER && TARGET_ARMARCH
+
+            if (!shouldOptimizeVirtualStubCall)
+            {
+                result = Ind(addr);
+            }
         }
     }
 

src/coreclr/src/jit/lsraarmarch.cpp

@@ -182,12 +182,10 @@ int LinearScan::BuildCall(GenTreeCall* call)
             ctrlExprCandidates = RBM_FASTTAILCALL_TARGET;
         }
     }
-#if defined(FEATURE_READYTORUN_COMPILER) && defined(TARGET_ARMARCH)
-    else if (call->IsR2RRelativeIndir())
+    else if (call->IsR2ROrVirtualStubRelativeIndir())
     {
         buildInternalIntRegisterDefForNode(call);
     }
-#endif // FEATURE_READYTORUN_COMPILER && TARGET_ARMARCH
 #ifdef TARGET_ARM
     else
     {