Fix a bug in lowerxarch related to merging Sse41.Insert chains

src/coreclr/jit/lowerxarch.cpp

@@ -2809,6 +2809,9 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
 
             if (comp->compOpportunisticallyDependsOn(InstructionSet_SSE41))
             {
+                assert(argCnt <= 4);
+                GenTree* insertedNodes[4];
+
                 for (N = 1; N < argCnt - 1; N++)
                 {
                     // We will be constructing the following parts:
@@ -2837,10 +2840,12 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
                     idx = comp->gtNewIconNode(N << 4, TYP_INT);
                     BlockRange().InsertAfter(tmp2, idx);
 
-                    tmp1 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, tmp2, idx, NI_SSE41_Insert, simdBaseJitType,
+                    tmp3 = comp->gtNewSimdHWIntrinsicNode(simdType, tmp1, tmp2, idx, NI_SSE41_Insert, simdBaseJitType,
                                                           simdSize);
-                    BlockRange().InsertAfter(idx, tmp1);
-                    LowerNode(tmp1);
+                    BlockRange().InsertAfter(idx, tmp3);
+
+                    insertedNodes[N] = tmp3;
+                    tmp1             = tmp3;
                 }
 
                 // We will be constructing the following parts:
@@ -2868,6 +2873,18 @@ GenTree* Lowering::LowerHWIntrinsicCreate(GenTreeHWIntrinsic* node)
                 BlockRange().InsertAfter(tmp2, idx);
 
                 node->ResetHWIntrinsicId(NI_SSE41_Insert, comp, tmp1, tmp2, idx);
+
+                for (N = 1; N < argCnt - 1; N++)
+                {
+                    // LowerNode for NI_SSE41_Insert specially handles zeros, constants, and certain mask values
+                    // to do the minimal number of operations and may merge together two neighboring inserts that
+                    // don't have any side effects between them. Because of this and because of the interdependence
+                    // of the inserts we've created above, we need to wait to lower the generated inserts until after
+                    // we've completed the chain.
+
+                    GenTree* insertedNode = insertedNodes[N];
+                    LowerNode(insertedNode);
+                }
                 break;
             }