vppinfra: vector shuffle cleanup

Type: refactor
Change-Id: I8b3fc2ce30df313467274a174c5ac6adbf296153
Signed-off-by: Damjan Marion <damarion@cisco.com>

6 files changed, 80 insertions, 53 deletions

diff --git a/src/examples/sample-plugin/sample/node.c b/src/examples/sample-plugin/sample/node.c
index a31c3e86e08..a9d8b66d788 100644
--- a/src/examples/sample-plugin/sample/node.c
+++ b/src/examples/sample-plugin/sample/node.c
@@ -291,7 +291,6 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
   sample_next_t next_index;
   u32 pkts_swapped = 0;
   /* Vector shuffle mask to swap src, dst */
-  u8x16 swapmac = { 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5, 12, 13, 14, 15 };
 
   from = vlib_frame_vector_args (frame);
   n_left_from = frame->n_vectors;
@@ -345,8 +344,10 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  src_dst0 = ((u8x16 *) en0)[0];
 	  src_dst1 = ((u8x16 *) en1)[0];
-	  src_dst0 = u8x16_shuffle (src_dst0, swapmac);
-	  src_dst1 = u8x16_shuffle (src_dst1, swapmac);
+	  src_dst0 = u8x16_shuffle (src_dst0, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3,
+				    4, 5, 12, 13, 14, 15);
+	  src_dst1 = u8x16_shuffle (src_dst1, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3,
+				    4, 5, 12, 13, 14, 15);
 	  ((u8x16 *) en0)[0] = src_dst0;
 	  ((u8x16 *) en1)[0] = src_dst1;
 
@@ -418,7 +419,8 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 	  en0 = vlib_buffer_get_current (b0);
 	  src_dst0 = ((u8x16 *) en0)[0];
-	  src_dst0 = u8x16_shuffle (src_dst0, swapmac);
+	  src_dst0 = u8x16_shuffle (src_dst0, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3,
+				    4, 5, 12, 13, 14, 15);
 	  ((u8x16 *) en0)[0] = src_dst0;
 
 	  sw_if_index0 = vnet_buffer (b0)->sw_if_index[VLIB_RX];
@@ -469,7 +471,6 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 #ifdef VERSION_3
 
-#define u8x16_shuffle __builtin_shuffle
 /* This would normally be a stack local, but since it's a constant... */
 static const u16 nexts[VLIB_FRAME_SIZE] = { 0 };
 
@@ -479,7 +480,6 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
   u32 n_left_from, *from;
   u32 pkts_swapped = 0;
   /* Vector shuffle mask to swap src, dst */
-  u8x16 swapmac = { 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5, 12, 13, 14, 15 };
   vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
   /* See comment below about sending all pkts to the same place... */
   u16 *next __attribute__ ((unused));
@@ -518,10 +518,14 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
       src_dst2 = ((u8x16 *) vlib_buffer_get_current (b[2]))[0];
       src_dst3 = ((u8x16 *) vlib_buffer_get_current (b[3]))[0];
 
-      src_dst0 = u8x16_shuffle (src_dst0, swapmac);
-      src_dst1 = u8x16_shuffle (src_dst1, swapmac);
-      src_dst2 = u8x16_shuffle (src_dst2, swapmac);
-      src_dst3 = u8x16_shuffle (src_dst3, swapmac);
+      src_dst0 = u8x16_shuffle (src_dst0, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15);
+      src_dst1 = u8x16_shuffle (src_dst1, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15);
+      src_dst2 = u8x16_shuffle (src_dst2, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15);
+      src_dst3 = u8x16_shuffle (src_dst3, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15);
 
       ((u8x16 *) vlib_buffer_get_current (b[0]))[0] = src_dst0;
       ((u8x16 *) vlib_buffer_get_current (b[1]))[0] = src_dst1;
@@ -552,7 +556,8 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
     {
       u8x16 src_dst0;
       src_dst0 = ((u8x16 *) vlib_buffer_get_current (b[0]))[0];
-      src_dst0 = u8x16_shuffle (src_dst0, swapmac);
+      src_dst0 = u8x16_shuffle (src_dst0, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5,
+				12, 13, 14, 15);
       ((u8x16 *) vlib_buffer_get_current (b[0]))[0] = src_dst0;
       vnet_buffer (b[0])->sw_if_index[VLIB_TX] =
 	vnet_buffer (b[0])->sw_if_index[VLIB_RX];
@@ -611,18 +616,14 @@ VLIB_NODE_FN (sample_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
 
 #ifdef VERSION_4
 
-#define u8x16_shuffle __builtin_shuffle
-
-static u8x16 swapmac =
-  { 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5, 12, 13, 14, 15 };
-
 /* Final stage in the pipeline, do the mac swap */
 static inline u32
 last_stage (vlib_main_t * vm, vlib_node_runtime_t * node, vlib_buffer_t * b)
 {
   u8x16 src_dst0;
   src_dst0 = ((u8x16 *) vlib_buffer_get_current (b))[0];
-  src_dst0 = u8x16_shuffle (src_dst0, swapmac);
+  src_dst0 = u8x16_shuffle (src_dst0, 6, 7, 8, 9, 10, 11, 0, 1, 2, 3, 4, 5, 12,
+			    13, 14, 15);
   ((u8x16 *) vlib_buffer_get_current (b))[0] = src_dst0;
   vnet_buffer (b)->sw_if_index[VLIB_TX] =
     vnet_buffer (b)->sw_if_index[VLIB_RX];
diff --git a/src/vppinfra/vector.h b/src/vppinfra/vector.h
index 88cf288cb26..49bc2976b3e 100644
--- a/src/vppinfra/vector.h
+++ b/src/vppinfra/vector.h
@@ -126,6 +126,68 @@ foreach_vec
 
 #undef _vector_size
 
+  /* _shuffle and _shuffle2 */
+#if defined(__GNUC__) && !defined(__clang__)
+#define __builtin_shufflevector(v1, v2, ...)                                  \
+  __builtin_shuffle ((v1), (v2), (__typeof__ (v1)){ __VA_ARGS__ })
+#endif
+
+#define u8x16_shuffle(v1, ...)                                                \
+  (u8x16) __builtin_shufflevector ((u8x16) (v1), (u8x16) (v1), __VA_ARGS__)
+#define u8x32_shuffle(v1, ...)                                                \
+  (u8x32) __builtin_shufflevector ((u8x32) (v1), (u8x32) (v1), __VA_ARGS__)
+#define u8x64_shuffle(v1, ...)                                                \
+  (u8x64) __builtin_shufflevector ((u8x64) (v1), (u8x64) (v1), __VA_ARGS__)
+
+#define u16x8_shuffle(v1, ...)                                                \
+  (u16x8) __builtin_shufflevector ((u16x8) (v1), (u16x8) (v1), __VA_ARGS__)
+#define u16x16_shuffle(v1, ...)                                               \
+  (u16x16) __builtin_shufflevector ((u16x16) (v1), (u16x16) (v1), __VA_ARGS__)
+#define u16x32_shuffle(v1, ...)                                               \
+  (u16u32) __builtin_shufflevector ((u16x32) (v1), (u16x32) (v1), __VA_ARGS__);
+
+#define u32x4_shuffle(v1, ...)                                                \
+  (u32x4) __builtin_shufflevector ((u32x4) (v1), (u32x4) (v1), __VA_ARGS__)
+#define u32x8_shuffle(v1, ...)                                                \
+  (u32x8) __builtin_shufflevector ((u32x8) (v1), (u32x8) (v1), __VA_ARGS__)
+#define u32x16_shuffle(v1, ...)                                               \
+  (u32x16) __builtin_shufflevector ((u32x16) (v1), (u32x16) (v1), __VA_ARGS__)
+
+#define u64x2_shuffle(v1, ...)                                                \
+  (u64x2) __builtin_shufflevector ((u64x2) (v1), (u64x2) (v1), __VA_ARGS__)
+#define u64x4_shuffle(v1, ...)                                                \
+  (u64x4) __builtin_shufflevector ((u64x4) (v1), (u64x4) (v1), __VA_ARGS__)
+#define u64x8_shuffle(v1, ...)                                                \
+  (u64x8) __builtin_shufflevector ((u64x8) (v1), (u64x8) (v1), __VA_ARGS__)
+
+#define u8x16_shuffle2(v1, v2, ...)                                           \
+  (u8x16) __builtin_shufflevector ((u8x16) (v1), (u8x16) (v2), __VA_ARGS__)
+#define u8x32_shuffle2(v1, v2, ...)                                           \
+  (u8x32) __builtin_shufflevector ((u8x32) (v1), (u8x32) (v2), __VA_ARGS__)
+#define u8x64_shuffle2(v1, v2, ...)                                           \
+  (u8x64) __builtin_shufflevector ((u8x64) (v1), (u8x64) (v2), __VA_ARGS__)
+
+#define u16x8_shuffle2(v1, v2, ...)                                           \
+  (u16x8) __builtin_shufflevector ((u16x8) (v1), (u16x8) (v2), __VA_ARGS__)
+#define u16x16_shuffle2(v1, v2, ...)                                          \
+  (u16x16) __builtin_shufflevector ((u16x16) (v1), (u16x16) (v2), __VA_ARGS__)
+#define u16x32_shuffle2(v1, v2, ...)                                          \
+  (u16u32) __builtin_shufflevector ((u16x32) (v1), (u16x32) (v2), __VA_ARGS__);
+
+#define u32x4_shuffle2(v1, v2, ...)                                           \
+  (u32x4) __builtin_shufflevector ((u32x4) (v1), (u32x4) (v2), __VA_ARGS__)
+#define u32x8_shuffle2(v1, v2, ...)                                           \
+  (u32x8) __builtin_shufflevector ((u32x8) (v1), (u32x8) (v2), __VA_ARGS__)
+#define u32x16_shuffle2(v1, v2, ...)                                          \
+  (u32x16) __builtin_shufflevector ((u32x16) (v1), (u32x16) (v2), __VA_ARGS__)
+
+#define u64x2_shuffle2(v1, v2, ...)                                           \
+  (u64x2) __builtin_shufflevector ((u64x2) (v1), (u64x2) (v2), __VA_ARGS__)
+#define u64x4_shuffle2(v1, v2, ...)                                           \
+  (u64x4) __builtin_shufflevector ((u64x4) (v1), (u64x4) (v2), __VA_ARGS__)
+#define u64x8_shuffle2(v1, v2, ...)                                           \
+  (u64x8) __builtin_shufflevector ((u64x8) (v1), (u64x8) (v2), __VA_ARGS__)
+
 #define VECTOR_WORD_TYPE(t) t##x
 #define VECTOR_WORD_TYPE_LEN(t) (sizeof (VECTOR_WORD_TYPE(t)) / sizeof (t))
 
diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h
index 7226c230e68..59857182a93 100644
--- a/src/vppinfra/vector_avx2.h
+++ b/src/vppinfra/vector_avx2.h
@@ -183,12 +183,6 @@ u16x16_byte_swap (u16x16 v)
   return (u16x16) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) swap);
 }
 
-static_always_inline u8x32
-u8x32_shuffle (u8x32 v, u8x32 m)
-{
-  return (u8x32) _mm256_shuffle_epi8 ((__m256i) v, (__m256i) m);
-}
-
 #define u8x32_align_right(a, b, imm) \
   (u8x32) _mm256_alignr_epi8 ((__m256i) a, (__m256i) b, imm)
 
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index 8acac2a3a9f..33f40ef7b5a 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -196,12 +196,6 @@ u8x64_reflect_u8x16 (u8x64 x)
   return (u8x64) _mm512_shuffle_epi8 ((__m512i) x, (__m512i) mask);
 }
 
-static_always_inline u8x64
-u8x64_shuffle (u8x64 v, u8x64 m)
-{
-  return (u8x64) _mm512_shuffle_epi8 ((__m512i) v, (__m512i) m);
-}
-
 #define u8x64_align_right(a, b, imm) \
   (u8x64) _mm512_alignr_epi8 ((__m512i) a, (__m512i) b, imm)
 
diff --git a/src/vppinfra/vector_neon.h b/src/vppinfra/vector_neon.h
index 80d7bda9f3c..e7b31259f73 100644
--- a/src/vppinfra/vector_neon.h
+++ b/src/vppinfra/vector_neon.h
@@ -129,12 +129,6 @@ u32x4_byte_swap (u32x4 v)
   return (u32x4) vrev32q_u8 ((u8x16) v);
 }
 
-static_always_inline u8x16
-u8x16_shuffle (u8x16 v, u8x16 m)
-{
-  return (u8x16) vqtbl1q_u8 (v, m);
-}
-
 static_always_inline u32x4
 u32x4_hadd (u32x4 v1, u32x4 v2)
 {
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index 7e75ad28710..35495d6658f 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -411,24 +411,6 @@ u32x4_sum_elts (u32x4 sum4)
   return sum4[0];
 }
 
-static_always_inline u8x16
-u8x16_shuffle (u8x16 v, u8x16 m)
-{
-  return (u8x16) _mm_shuffle_epi8 ((__m128i) v, (__m128i) m);
-}
-
-static_always_inline u32x4
-u32x4_shuffle (u32x4 v, const int a, const int b, const int c, const int d)
-{
-#if defined(__clang__) || !__OPTIMIZE__
-  u32x4 r = { v[a], v[b], v[c], v[d] };
-  return r;
-#else
-  return (u32x4) _mm_shuffle_epi32 ((__m128i) v,
-				    a | b << 2 | c << 4 | d << 6);
-#endif
-}
-
 /* _from_ */
 /* *INDENT-OFF* */
 #define _(f,t,i) \