Improve the handling of ToScalar and GetElement(0)

src/coreclr/jit/codegenxarch.cpp

@@ -5405,6 +5405,9 @@ void CodeGen::genCodeForStoreInd(GenTreeStoreInd* tree)
 
                     switch (intrinsicId)
                     {
+                        case NI_Vector128_ToScalar:
+                        case NI_Vector256_ToScalar:
+                        case NI_Vector512_ToScalar:
                         case NI_SSE2_ConvertToInt32:
                         case NI_SSE2_ConvertToUInt32:
                         case NI_SSE2_X64_ConvertToInt64:

src/coreclr/jit/decomposelongs.cpp

@@ -1702,11 +1702,15 @@ GenTree* DecomposeLongs::DecomposeHWIntrinsic(LIR::Use& use)
         case NI_Vector128_GetElement:
         case NI_Vector256_GetElement:
         case NI_Vector512_GetElement:
+        {
             return DecomposeHWIntrinsicGetElement(use, hwintrinsicTree);
+        }
 
         default:
+        {
             noway_assert(!"unexpected GT_HWINTRINSIC node in long decomposition");
             break;
+        }
     }
 
     return nullptr;

src/coreclr/jit/gentree.cpp

@@ -19067,6 +19067,9 @@ bool GenTree::isContainableHWIntrinsic() const
         }
 
         case NI_Vector128_GetElement:
+        case NI_Vector128_ToScalar:
+        case NI_Vector256_ToScalar:
+        case NI_Vector512_ToScalar:
         case NI_SSE2_ConvertToInt32:
         case NI_SSE2_ConvertToUInt32:
         case NI_SSE2_X64_ConvertToInt64:
@@ -22057,27 +22060,56 @@ GenTree* Compiler::gtNewSimdGetElementNode(
     assert(varTypeIsArithmetic(simdBaseType));
 
 #if defined(TARGET_XARCH)
+    bool useToScalar = op2->IsIntegralConst(0);
+
+#if defined(TARGET_X86)
+    // We handle decomposition via GetElement for simplicity
+    useToScalar &= !varTypeIsLong(simdBaseType);
+#endif // TARGET_X86
+
+    if (useToScalar)
+    {
+        intrinsicId = NI_Vector128_ToScalar;
+
+        if (simdSize == 64)
+        {
+            intrinsicId = NI_Vector512_ToScalar;
+        }
+        else if (simdSize == 32)
+        {
+            intrinsicId = NI_Vector256_ToScalar;
+        }
+
+        return gtNewSimdHWIntrinsicNode(type, op1, intrinsicId, simdBaseJitType, simdSize);
+    }
+
     switch (simdBaseType)
     {
-        // Using software fallback if simdBaseType is not supported by hardware
         case TYP_BYTE:
         case TYP_UBYTE:
         case TYP_INT:
         case TYP_UINT:
         case TYP_LONG:
         case TYP_ULONG:
+        {
+            // Using software fallback if simdBaseType is not supported by hardware
             assert(compIsaSupportedDebugOnly(InstructionSet_SSE41));
             break;
+        }
 
         case TYP_DOUBLE:
         case TYP_FLOAT:
         case TYP_SHORT:
         case TYP_USHORT:
+        {
             assert(compIsaSupportedDebugOnly(InstructionSet_SSE2));
             break;
+        }
 
         default:
+        {
             unreached();
+        }
     }
 
     if (simdSize == 64)
@@ -22089,6 +22121,18 @@ GenTree* Compiler::gtNewSimdGetElementNode(
         intrinsicId = NI_Vector256_GetElement;
     }
 #elif defined(TARGET_ARM64)
+    if (op2->IsIntegralConst(0))
+    {
+        intrinsicId = NI_Vector128_ToScalar;
+
+        if (simdSize == 8)
+        {
+            intrinsicId = NI_Vector64_ToScalar;
+        }
+
+        return gtNewSimdHWIntrinsicNode(type, op1, intrinsicId, simdBaseJitType, simdSize);
+    }
+
     if (simdSize == 8)
     {
         intrinsicId = NI_Vector64_GetElement;

src/coreclr/jit/hwintrinsicxarch.cpp

@@ -1678,34 +1678,48 @@ GenTree* Compiler::impSpecialIntrinsic(NamedIntrinsic        intrinsic,
         {
             assert(sig->numArgs == 2);
 
+            op2 = impStackTop(0).val;
+
             switch (simdBaseType)
             {
-                // Using software fallback if simdBaseType is not supported by hardware
                 case TYP_BYTE:
                 case TYP_UBYTE:
                 case TYP_INT:
                 case TYP_UINT:
                 case TYP_LONG:
                 case TYP_ULONG:
-                    if (!compExactlyDependsOn(InstructionSet_SSE41))
+                {
+                    bool useToScalar = op2->IsIntegralConst(0);
+
+#if defined(TARGET_X86)
+                    useToScalar &= !varTypeIsLong(simdBaseType);
+#endif // TARGET_X86
+
+                    if (!useToScalar && !compExactlyDependsOn(InstructionSet_SSE41))
                     {
+                        // Using software fallback if simdBaseType is not supported by hardware
                         return nullptr;
                     }
                     break;
+                }
 
                 case TYP_DOUBLE:
                 case TYP_FLOAT:
                 case TYP_SHORT:
                 case TYP_USHORT:
+                {
                     // short/ushort/float/double is supported by SSE2
                     break;
+                }
 
                 default:
+                {
                     unreached();
+                }
             }
 
-            GenTree* op2 = impPopStack().val;
-            GenTree* op1 = impSIMDPopStack();
+            impPopStack();
+            op1 = impSIMDPopStack();
 
             retNode = gtNewSimdGetElementNode(retType, op1, op2, simdBaseJitType, simdSize);
             break;
@@ -2543,16 +2557,18 @@ GenTree* Compiler::impSpecialIntrinsic(NamedIntrinsic        intrinsic,
         {
             assert(sig->numArgs == 1);
 
+            op1 = impSIMDPopStack();
+
 #if defined(TARGET_X86)
             if (varTypeIsLong(simdBaseType))
             {
-                // TODO-XARCH-CQ: It may be beneficial to decompose this operation
+                // Create a GetElement node which handles decomposition
+                op2     = gtNewIconNode(0);
+                retNode = gtNewSimdGetElementNode(retType, op1, op2, simdBaseJitType, simdSize);
                 break;
             }
 #endif // TARGET_X86
 
-            // TODO-XARCH-CQ: It may be beneficial to import this as GetElement(0)
-            op1     = impSIMDPopStack();
             retNode = gtNewSimdHWIntrinsicNode(retType, op1, intrinsic, simdBaseJitType, simdSize);
             break;
         }