Fix a bug around Vector4.Distance and Sse41.Insert lowering (#81725)

* Adding a regression test for #81585

* Fixing a bug around Sse41.Insert lowering

* Ensure that Distance/DistanceSquared are correctly imported

* Account for another case around Sse41.Insert chain folding

Author: tannergooding

src/coreclr/jit/lowerxarch.cpp

@@ -1314,10 +1314,14 @@ GenTree* Lowering::LowerHWIntrinsic(GenTreeHWIntrinsic* node)
                 // the zmask from op1. We expect that op2 has already been
                 // lowered and therefore the containment checks have happened
 
+                // Since this is a newer operation, we need to account for
+                // the possibility of `op1Intrinsic` zeroing the same element
+                // we're setting here.
+
                 assert(op1Intrinsic->Op(2)->isContained());
 
                 ssize_t op1Ival = op1Idx->AsIntConCommon()->IconValue();
-                ival |= (op1Ival & 0x0F);
+                ival |= ((op1Ival & 0x0F) & ~(1 << count_d));
                 op3->AsIntConCommon()->SetIconValue(ival);
 
                 // Then we'll just carry the original non-zero input and
@@ -1335,6 +1339,8 @@ GenTree* Lowering::LowerHWIntrinsic(GenTreeHWIntrinsic* node)
                 // account, we can basically do the same thing here by merging this
                 // zmask into the ival from op1.
 
+                // Since this is a later op, direct merging is safe
+
                 ssize_t op1Ival = op1Idx->AsIntConCommon()->IconValue();
                 ival            = op1Ival | zmask;
                 op3->AsIntConCommon()->SetIconValue(ival);

src/coreclr/jit/simdashwintrinsic.cpp

@@ -946,39 +946,6 @@ GenTree* Compiler::impSimdAsHWIntrinsicSpecial(NamedIntrinsic       intrinsic,
                                               /* isSimdAsHWIntrinsic */ true);
                 }
 
-                case NI_Vector2_Distance:
-                case NI_Vector3_Distance:
-                case NI_Vector4_Distance:
-                {
-                    op1 = gtNewSimdBinOpNode(GT_SUB, retType, op1, op2, simdBaseJitType, simdSize,
-                                             /* isSimdAsHWIntrinsic */ true);
-
-                    GenTree* clonedOp1;
-                    op1 = impCloneExpr(op1, &clonedOp1, NO_CLASS_HANDLE, CHECK_SPILL_ALL,
-                                       nullptr DEBUGARG("Clone diff for vector distance"));
-
-                    op1 = gtNewSimdDotProdNode(retType, op1, clonedOp1, simdBaseJitType, simdSize,
-                                               /* isSimdAsHWIntrinsic */ true);
-
-                    return new (this, GT_INTRINSIC)
-                        GenTreeIntrinsic(simdBaseType, op1, NI_System_Math_Sqrt, NO_METHOD_HANDLE);
-                }
-
-                case NI_Vector2_DistanceSquared:
-                case NI_Vector3_DistanceSquared:
-                case NI_Vector4_DistanceSquared:
-                {
-                    op1 = gtNewSimdBinOpNode(GT_SUB, retType, op1, op2, simdBaseJitType, simdSize,
-                                             /* isSimdAsHWIntrinsic */ true);
-
-                    GenTree* clonedOp1;
-                    op1 = impCloneExpr(op1, &clonedOp1, NO_CLASS_HANDLE, CHECK_SPILL_ALL,
-                                       nullptr DEBUGARG("Clone diff for vector distance squared"));
-
-                    return gtNewSimdDotProdNode(retType, op1, clonedOp1, simdBaseJitType, simdSize,
-                                                /* isSimdAsHWIntrinsic */ true);
-                }
-
                 case NI_VectorT128_Floor:
 #if defined(TARGET_XARCH)
                 case NI_VectorT256_Floor:
@@ -1382,6 +1349,39 @@ GenTree* Compiler::impSimdAsHWIntrinsicSpecial(NamedIntrinsic       intrinsic,
                     break;
                 }
 
+                case NI_Vector2_Distance:
+                case NI_Vector3_Distance:
+                case NI_Vector4_Distance:
+                {
+                    op1 = gtNewSimdBinOpNode(GT_SUB, simdType, op1, op2, simdBaseJitType, simdSize,
+                                             /* isSimdAsHWIntrinsic */ true);
+
+                    GenTree* clonedOp1;
+                    op1 = impCloneExpr(op1, &clonedOp1, NO_CLASS_HANDLE, CHECK_SPILL_ALL,
+                                       nullptr DEBUGARG("Clone diff for vector distance"));
+
+                    op1 = gtNewSimdDotProdNode(retType, op1, clonedOp1, simdBaseJitType, simdSize,
+                                               /* isSimdAsHWIntrinsic */ true);
+
+                    return new (this, GT_INTRINSIC)
+                        GenTreeIntrinsic(retType, op1, NI_System_Math_Sqrt, NO_METHOD_HANDLE);
+                }
+
+                case NI_Vector2_DistanceSquared:
+                case NI_Vector3_DistanceSquared:
+                case NI_Vector4_DistanceSquared:
+                {
+                    op1 = gtNewSimdBinOpNode(GT_SUB, simdType, op1, op2, simdBaseJitType, simdSize,
+                                             /* isSimdAsHWIntrinsic */ true);
+
+                    GenTree* clonedOp1;
+                    op1 = impCloneExpr(op1, &clonedOp1, NO_CLASS_HANDLE, CHECK_SPILL_ALL,
+                                       nullptr DEBUGARG("Clone diff for vector distance squared"));
+
+                    return gtNewSimdDotProdNode(retType, op1, clonedOp1, simdBaseJitType, simdSize,
+                                                /* isSimdAsHWIntrinsic */ true);
+                }
+
                 case NI_Quaternion_Divide:
                 case NI_Vector2_Divide:
                 case NI_Vector2_op_Division: