+#ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
+/* SSE version of FDIR mark extraction for 4 32B descriptors at a time */
+static inline __m128i
+descs_to_fdir_32b(volatile union i40e_rx_desc *rxdp, struct rte_mbuf **rx_pkt)
+{
+ /* 32B descriptors: Load 2nd half of descriptors for FDIR ID data */
+ __m128i desc0_qw23, desc1_qw23, desc2_qw23, desc3_qw23;
+ desc0_qw23 = _mm_loadu_si128((__m128i *)&(rxdp + 0)->wb.qword2);
+ desc1_qw23 = _mm_loadu_si128((__m128i *)&(rxdp + 1)->wb.qword2);
+ desc2_qw23 = _mm_loadu_si128((__m128i *)&(rxdp + 2)->wb.qword2);
+ desc3_qw23 = _mm_loadu_si128((__m128i *)&(rxdp + 3)->wb.qword2);
+
+ /* FDIR ID data: move last u32 of each desc to 4 u32 lanes */
+ __m128i v_unpack_01, v_unpack_23;
+ v_unpack_01 = _mm_unpackhi_epi32(desc0_qw23, desc1_qw23);
+ v_unpack_23 = _mm_unpackhi_epi32(desc2_qw23, desc3_qw23);
+ __m128i v_fdir_ids = _mm_unpackhi_epi64(v_unpack_01, v_unpack_23);
+
+ /* Extended Status: extract from each lower 32 bits, to u32 lanes */
+ v_unpack_01 = _mm_unpacklo_epi32(desc0_qw23, desc1_qw23);
+ v_unpack_23 = _mm_unpacklo_epi32(desc2_qw23, desc3_qw23);
+ __m128i v_flt_status = _mm_unpacklo_epi64(v_unpack_01, v_unpack_23);
+
+ /* Shift u32 left and right to "mask away" bits not required.
+ * Data required is 4:5 (zero based), so left shift by 26 (32-6)
+ * and then right shift by 30 (32 - 2 bits required).
+ */
+ v_flt_status = _mm_slli_epi32(v_flt_status, 26);
+ v_flt_status = _mm_srli_epi32(v_flt_status, 30);
+
+ /* Generate constant 1 in all u32 lanes and compare */
+ RTE_BUILD_BUG_ON(I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID != 1);
+ __m128i v_zeros = _mm_setzero_si128();
+ __m128i v_ffff = _mm_cmpeq_epi32(v_zeros, v_zeros);
+ __m128i v_u32_one = _mm_srli_epi32(v_ffff, 31);
+
+ /* per desc mask, bits set if FDIR ID is valid */
+ __m128i v_fd_id_mask = _mm_cmpeq_epi32(v_flt_status, v_u32_one);
+
+ /* Mask ID data to zero if the FD_ID bit not set in desc */
+ v_fdir_ids = _mm_and_si128(v_fdir_ids, v_fd_id_mask);
+
+ /* Extract and store as u32. No advantage to combining into SSE
+ * stores, there are no surrounding stores to around fdir.hi
+ */
+ rx_pkt[0]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 0);
+ rx_pkt[1]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 1);
+ rx_pkt[2]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 2);
+ rx_pkt[3]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 3);
+
+ /* convert fdir_id_mask into a single bit, then shift as required for
+ * correct location in the mbuf->olflags
+ */
+ const uint32_t FDIR_ID_BIT_SHIFT = 13;
+ RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << FDIR_ID_BIT_SHIFT));
+ v_fd_id_mask = _mm_srli_epi32(v_fd_id_mask, 31);
+ v_fd_id_mask = _mm_slli_epi32(v_fd_id_mask, FDIR_ID_BIT_SHIFT);
+
+ /* The returned value must be combined into each mbuf. This is already
+ * being done for RSS and VLAN mbuf olflags, so return bits to OR in.
+ */
+ return v_fd_id_mask;
+}
+
+#else /* 32 or 16B FDIR ID handling */
+
+/* Handle 16B descriptor FDIR ID flag setting based on FLM. See scalar driver
+ * for scalar implementation of the same functionality.
+ */
+static inline __m128i
+descs_to_fdir_16b(__m128i fltstat, __m128i descs[4], struct rte_mbuf **rx_pkt)
+{
+ /* unpack filter-status data from descriptors */
+ __m128i v_tmp_01 = _mm_unpacklo_epi32(descs[0], descs[1]);
+ __m128i v_tmp_23 = _mm_unpacklo_epi32(descs[2], descs[3]);
+ __m128i v_fdir_ids = _mm_unpackhi_epi64(v_tmp_01, v_tmp_23);
+
+ /* Generate one bit in each u32 lane */
+ __m128i v_zeros = _mm_setzero_si128();
+ __m128i v_ffff = _mm_cmpeq_epi32(v_zeros, v_zeros);
+ __m128i v_111_mask = _mm_srli_epi32(v_ffff, 29);
+ __m128i v_11_mask = _mm_srli_epi32(v_ffff, 30);
+
+ /* Top lane ones mask for FDIR isolation */
+ __m128i v_desc_fdir_mask = _mm_insert_epi32(v_zeros, UINT32_MAX, 1);
+
+ /* Compare and mask away FDIR ID data if bit not set */
+ __m128i v_u32_bits = _mm_and_si128(v_111_mask, fltstat);
+ __m128i v_fdir_id_mask = _mm_cmpeq_epi32(v_u32_bits, v_11_mask);
+ v_fdir_ids = _mm_and_si128(v_fdir_id_mask, v_fdir_ids);
+
+ /* Store data to fdir.hi in mbuf */
+ rx_pkt[0]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 0);
+ rx_pkt[1]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 1);
+ rx_pkt[2]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 2);
+ rx_pkt[3]->hash.fdir.hi = _mm_extract_epi32(v_fdir_ids, 3);
+
+ /* Move fdir_id_mask to correct lane, blend RSS to zero on hits */
+ __m128i v_desc3_shift = _mm_alignr_epi8(v_zeros, v_fdir_id_mask, 8);
+ __m128i v_desc3_mask = _mm_and_si128(v_desc_fdir_mask, v_desc3_shift);
+ descs[3] = _mm_blendv_epi8(descs[3], _mm_setzero_si128(), v_desc3_mask);
+
+ __m128i v_desc2_shift = _mm_alignr_epi8(v_zeros, v_fdir_id_mask, 4);
+ __m128i v_desc2_mask = _mm_and_si128(v_desc_fdir_mask, v_desc2_shift);
+ descs[2] = _mm_blendv_epi8(descs[2], _mm_setzero_si128(), v_desc2_mask);
+
+ __m128i v_desc1_shift = v_fdir_id_mask;
+ __m128i v_desc1_mask = _mm_and_si128(v_desc_fdir_mask, v_desc1_shift);
+ descs[1] = _mm_blendv_epi8(descs[1], _mm_setzero_si128(), v_desc1_mask);
+
+ __m128i v_desc0_shift = _mm_alignr_epi8(v_fdir_id_mask, v_zeros, 12);
+ __m128i v_desc0_mask = _mm_and_si128(v_desc_fdir_mask, v_desc0_shift);
+ descs[0] = _mm_blendv_epi8(descs[0], _mm_setzero_si128(), v_desc0_mask);
+
+ /* Shift to 1 or 0 bit per u32 lane, then to PKT_RX_FDIR_ID offset */
+ const uint32_t FDIR_ID_BIT_SHIFT = 13;
+ RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << FDIR_ID_BIT_SHIFT));
+ __m128i v_mask_one_bit = _mm_srli_epi32(v_fdir_id_mask, 31);
+ return _mm_slli_epi32(v_mask_one_bit, FDIR_ID_BIT_SHIFT);
+}
+#endif
+