vector functions cleanup and improvements

Remove functions which have native C equivalent (i.e. _is_equal can be
replaced with ==, _add with +)
Add SSE4.2, AVX-512 implementations of splat, load_unaligned, store_unaligned,
is_all_zero, is_equal, is_all_equal

Change-Id: Ie80b0e482e7a76248ad79399c2576468532354cd
Signed-off-by: Damjan Marion <damarion@cisco.com>

8 files changed, 99 insertions, 96 deletions

diff --git a/src/vnet/ip/ip4_source_and_port_range_check.c b/src/vnet/ip/ip4_source_and_port_range_check.c
index 06e6e7c6b63..2889a899678 100644
--- a/src/vnet/ip/ip4_source_and_port_range_check.c
+++ b/src/vnet/ip/ip4_source_and_port_range_check.c
@@ -133,9 +133,8 @@ check_adj_port_range_x1 (const protocol_port_range_dpo_t * ppr_dpo,
 	u16x8_sub_saturate (ppr_dpo->blocks[i].low.as_u16x8, key.as_u16x8);
       diff2.as_u16x8 =
 	u16x8_sub_saturate (ppr_dpo->blocks[i].hi.as_u16x8, key.as_u16x8);
-      sum.as_u16x8 = u16x8_add (diff1.as_u16x8, diff2.as_u16x8);
-      sum_equal_diff2.as_u16x8 =
-	u16x8_is_equal (sum.as_u16x8, diff2.as_u16x8);
+      sum.as_u16x8 = diff1.as_u16x8 + diff2.as_u16x8;
+      sum_equal_diff2.as_u16x8 = (sum.as_u16x8 == diff2.as_u16x8);
       sum_nonzero = ~u16x8_zero_byte_mask (sum.as_u16x8);
       sum_equal = ~u16x8_zero_byte_mask (sum_equal_diff2.as_u16x8);
       winner_mask = sum_nonzero & sum_equal;
diff --git a/src/vppinfra/mheap.c b/src/vppinfra/mheap.c
index c703545954d..4d27d419e64 100644
--- a/src/vppinfra/mheap.c
+++ b/src/vppinfra/mheap.c
@@ -311,7 +311,7 @@ mheap_small_object_cache_mask (mheap_small_object_cache_t * c, uword bin)
 
   ASSERT (bin < 256);
 
-#define _(i) ((uword) u8x16_compare_byte_mask (u8x16_is_equal (b, c->bins.as_u8x16[i])) << (uword) ((i)*16))
+#define _(i) ((uword) u8x16_compare_byte_mask ((b == c->bins.as_u8x16[i])) << (uword) ((i)*16))
   mask = _(0) | _(1);
   if (BITS (uword) > 32)
     mask |= _(2) | _(3);
diff --git a/src/vppinfra/pfhash.h b/src/vppinfra/pfhash.h
index e054c668f3b..2884fa81cf9 100644
--- a/src/vppinfra/pfhash.h
+++ b/src/vppinfra/pfhash.h
@@ -249,8 +249,8 @@ pfhash_search_kv_4 (pfhash_t * p, u32 bucket_contents, u32 * key)
 
   vector_key = u32x4_splat (key[0]);
 
-  is_equal[0] = u32x4_is_equal (kv->kb.k_u32x4[0], vector_key);
-  is_equal[1] = u32x4_is_equal (kv->kb.k_u32x4[1], vector_key);
+  is_equal[0] = (kv->kb.k_u32x4[0] == vector_key);
+  is_equal[1] = (kv->kb.k_u32x4[1] == vector_key);
   zbm[0] = ~u32x4_zero_byte_mask (is_equal[0]) & 0xFFFF;
   zbm[1] = ~u32x4_zero_byte_mask (is_equal[1]) & 0xFFFF;
 
diff --git a/src/vppinfra/vector.h b/src/vppinfra/vector.h
index fcff5e79d95..2157ab7d1d0 100644
--- a/src/vppinfra/vector.h
+++ b/src/vppinfra/vector.h
@@ -157,47 +157,6 @@ typedef u64 u64x _vector_size (8);
 #define VECTOR_WORD_TYPE(t) t##x
 #define VECTOR_WORD_TYPE_LEN(t) (sizeof (VECTOR_WORD_TYPE(t)) / sizeof (t))
 
-/* this series of macros generate _is_equal, _is_greater, _is_zero, _add
-   and _sub inline funcitons for each vector type */
-#define _(t, s, c) \
-  static_always_inline t##s##x##c			\
-t##s##x##c##_is_equal (t##s##x##c v1, t##s##x##c v2)	\
-{ return (v1 == v2); }					\
-							\
-static_always_inline t##s##x##c				\
-t##s##x##c##_is_greater (t##s##x##c v1, t##s##x##c v2)	\
-{ return (v1 > v2); }					\
-							\
-static_always_inline t##s##x##c				\
-t##s##x##c##_is_zero (t##s##x##c v1)			\
-{ t##s##x##c z = {0}; return (v1 == z); }		\
-							\
-static_always_inline t##s##x##c				\
-t##s##x##c##_add (t##s##x##c v1, t##s##x##c v2)		\
-{ return (v1 + v2); }					\
-							\
-static_always_inline t##s##x##c				\
-t##s##x##c##_sub (t##s##x##c v1, t##s##x##c v2)		\
-{ return (v1 - v2); }
-  foreach_vec
-#undef _
-
-/* this macro generate _splat inline functions for each scalar vector type */
-#define _(t, s, c) \
-  static_always_inline t##s##x##c			\
-t##s##x##c##_splat (t##s x)				\
-{							\
-    t##s##x##c r;					\
-    int i;						\
-							\
-    for (i = 0; i < c; i++)				\
-      r[i] = x;						\
-							\
-    return r;						\
-}
-  foreach_vec128i foreach_vec128u
-#undef _
-
 #if defined (__SSE4_2__) && __GNUC__ >= 4
 #include <vppinfra/vector_sse42.h>
 #endif
@@ -222,6 +181,24 @@ t##s##x##c##_splat (t##s x)				\
 #include <vppinfra/vector_funcs.h>
 #endif
 
+/* this macro generate _splat inline functions for each scalar vector type */
+#ifndef CLIB_VEC128_SPLAT_DEFINED
+#define _(t, s, c) \
+  static_always_inline t##s##x##c			\
+t##s##x##c##_splat (t##s x)				\
+{							\
+    t##s##x##c r;					\
+    int i;						\
+							\
+    for (i = 0; i < c; i++)				\
+      r[i] = x;						\
+							\
+    return r;						\
+}
+  foreach_vec128i foreach_vec128u
+#undef _
+#endif
+
 /* *INDENT-ON* */
 
 #endif /* included_clib_vector_h */
diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h
index 3f0b397b828..1fb41dfd7df 100644
--- a/src/vppinfra/vector_avx2.h
+++ b/src/vppinfra/vector_avx2.h
@@ -19,6 +19,7 @@
 #include <vppinfra/clib.h>
 #include <x86intrin.h>
 
+/* *INDENT-OFF* */
 #define foreach_avx2_vec256i \
   _(i,8,32,epi8) _(i,16,16,epi16) _(i,32,8,epi32)  _(i,64,4,epi64x)
 #define foreach_avx2_vec256u \
@@ -26,7 +27,8 @@
 #define foreach_avx2_vec256f \
   _(f,32,8,ps) _(f,64,4,pd)
 
-/* splat, load_unaligned, store_unaligned, is_all_zero, is_all_equal */
+/* splat, load_unaligned, store_unaligned, is_all_zero, is_equal,
+   is_all_equal */
 #define _(t, s, c, i) \
 static_always_inline t##s##x##c						\
 t##s##x##c##_splat (t##s x)						\
@@ -45,13 +47,18 @@ t##s##x##c##_is_all_zero (t##s##x##c x)					\
 { return _mm256_testz_si256 ((__m256i) x, (__m256i) x); }		\
 \
 static_always_inline int						\
-t##s##x##c##_is_all_equal (t##s##x##c v, t##s x)			\
-{ return t##s##x##c##_is_all_zero (v != t##s##x##c##_splat (x)); };	\
+t##s##x##c##_is_equal (t##s##x##c x, t##s##x##c y)			\
+{ return _mm256_testc_si256 ((__m256i) x, (__m256i) y); }		\
 \
+static_always_inline int						\
+t##s##x##c##_is_all_equal (t##s##x##c v, t##s x)			\
+{ return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); };		\
 
 foreach_avx2_vec256i foreach_avx2_vec256u
 #undef _
-  always_inline u32x8
+/* *INDENT-ON* */
+
+always_inline u32x8
 u32x8_permute (u32x8 v, u32x8 idx)
 {
   return (u32x8) _mm256_permutevar8x32_epi32 ((__m256i) v, (__m256i) idx);
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index c1b7c42a260..c2903e2aa1a 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -19,6 +19,7 @@
 #include <vppinfra/clib.h>
 #include <x86intrin.h>
 
+/* *INDENT-OFF* */
 #define foreach_avx512_vec512i \
   _(i,8,64,epi8) _(i,16,32,epi16) _(i,32,16,epi32)  _(i,64,8,epi64)
 #define foreach_avx512_vec512u \
@@ -26,8 +27,8 @@
 #define foreach_avx512_vec512f \
   _(f,32,8,ps) _(f,64,4,pd)
 
-/* splat, load_unaligned, store_unaligned */
-/* *INDENT-OFF* */
+/* splat, load_unaligned, store_unaligned, is_all_zero, is_equal,
+   is_all_equal */
 #define _(t, s, c, i) \
 static_always_inline t##s##x##c						\
 t##s##x##c##_splat (t##s x)						\
@@ -41,6 +42,17 @@ static_always_inline void						\
 t##s##x##c##_store_unaligned (t##s##x##c v, void *p)			\
 { _mm512_storeu_si512 ((__m512i *) p, (__m512i) v); }			\
 \
+static_always_inline int						\
+t##s##x##c##_is_all_zero (t##s##x##c v)					\
+{ return (_mm512_test_epi64_mask ((__m512i) v, (__m512i) v) == 0); }	\
+\
+static_always_inline int						\
+t##s##x##c##_is_equal (t##s##x##c a, t##s##x##c b)			\
+{ return t##s##x##c##_is_all_zero (a ^b); }				\
+\
+static_always_inline int						\
+t##s##x##c##_is_all_equal (t##s##x##c v, t##s x)			\
+{ return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); }		\
 
 foreach_avx512_vec512i foreach_avx512_vec512u
 #undef _
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index 0e334c828b6..053826d92bc 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -41,6 +41,48 @@
 #include <vppinfra/error_bootstrap.h>	/* for ASSERT */
 #include <x86intrin.h>
 
+/* *INDENT-OFF* */
+#define foreach_sse42_vec128i \
+  _(i,8,16,epi8) _(i,16,8,epi16) _(i,32,4,epi32)  _(i,64,2,epi64x)
+#define foreach_sse42_vec128u \
+  _(u,8,16,epi8) _(u,16,8,epi16) _(u,32,4,epi32)  _(u,64,2,epi64x)
+#define foreach_sse42_vec128f \
+  _(f,32,4,ps) _(f,64,2,pd)
+
+/* splat, load_unaligned, store_unaligned, is_all_zero, is_equal,
+   is_all_equal */
+#define _(t, s, c, i) \
+static_always_inline t##s##x##c						\
+t##s##x##c##_splat (t##s x)						\
+{ return (t##s##x##c) _mm_set1_##i (x); }				\
+\
+static_always_inline t##s##x##c						\
+t##s##x##c##_load_unaligned (void *p)					\
+{ return (t##s##x##c) _mm_loadu_si128 (p); }				\
+\
+static_always_inline void						\
+t##s##x##c##_store_unaligned (t##s##x##c v, void *p)			\
+{ _mm_storeu_si128 ((__m128i *) p, (__m128i) v); }			\
+\
+static_always_inline int						\
+t##s##x##c##_is_all_zero (t##s##x##c x)					\
+{ return _mm_testz_si128 ((__m128i) x, (__m128i) x); }			\
+\
+static_always_inline int						\
+t##s##x##c##_is_equal (t##s##x##c x, t##s##x##c y)			\
+{ return _mm_testc_si128 ((__m128i) x, (__m128i) y); }			\
+\
+static_always_inline int						\
+t##s##x##c##_is_all_equal (t##s##x##c v, t##s x)			\
+{ return t##s##x##c##_is_equal (v, t##s##x##c##_splat (x)); };		\
+
+foreach_sse42_vec128i foreach_sse42_vec128u
+#undef _
+/* *INDENT-ON* */
+
+#define CLIB_VEC128_SPLAT_DEFINED
+#define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
+
 /* 128 bit interleaves. */
 always_inline u8x16
 u8x16_interleave_hi (u8x16 a, u8x16 b)
@@ -197,16 +239,6 @@ u64x2_write_hi (u64x2 x, u64 * a)
 }
 #endif
 
-/* Unaligned loads/stores. */
-
-#define _(t)						\
-  always_inline void t##_store_unaligned (t x, void * a)	\
-  { _mm_storeu_si128 ((__m128i *) a, (__m128i) x); }	\
-  always_inline t t##_load_unaligned (void * a)		\
-  { return (t) _mm_loadu_si128 ((__m128i *) a); }
-
-_(u8x16) _(u16x8) _(u32x4) _(u64x2) _(i8x16) _(i16x8) _(i32x4) _(i64x2)
-#undef _
 #define _signed_binop(n,m,f,g)                                         \
   /* Unsigned */                                                       \
   always_inline u##n##x##m                                             \
@@ -218,7 +250,7 @@ _(u8x16) _(u16x8) _(u32x4) _(u64x2) _(i8x16) _(i16x8) _(i32x4) _(i64x2)
   i##n##x##m##_##f (i##n##x##m x, i##n##x##m y)                        \
   { return (i##n##x##m) _mm_##g##n ((__m128i) x, (__m128i) y); }
 /* Addition/subtraction with saturation. */
-  _signed_binop (8, 16, add_saturate, adds_epu)
+_signed_binop (8, 16, add_saturate, adds_epu)
 _signed_binop (16, 8, add_saturate, adds_epu)
 _signed_binop (8, 16, sub_saturate, subs_epu)
 _signed_binop (16, 8, sub_saturate, subs_epu)
@@ -403,30 +435,6 @@ _(u64, 2, right, left);
 #undef _
 #endif
 
-always_inline int
-u8x16_is_all_zero (u8x16 x)
-{
-  return _mm_testz_si128 ((__m128i) x, (__m128i) x);
-}
-
-always_inline int
-u16x8_is_all_zero (u16x8 x)
-{
-  return _mm_testz_si128 ((__m128i) x, (__m128i) x);
-}
-
-always_inline int
-u32x4_is_all_zero (u32x4 x)
-{
-  return _mm_testz_si128 ((__m128i) x, (__m128i) x);
-}
-
-always_inline int
-u64x2_is_all_zero (u64x2 x)
-{
-  return _mm_testz_si128 ((__m128i) x, (__m128i) x);
-}
-
 #define u32x4_select(A,MASK)						\
 ({									\
   u32x4 _x, _y;								\
@@ -495,21 +503,21 @@ always_inline u32
 u8x16_zero_byte_mask (u8x16 x)
 {
   u8x16 zero = { 0 };
-  return u8x16_compare_byte_mask (u8x16_is_equal (x, zero));
+  return u8x16_compare_byte_mask (x == zero);
 }
 
 always_inline u32
 u16x8_zero_byte_mask (u16x8 x)
 {
   u16x8 zero = { 0 };
-  return u8x16_compare_byte_mask ((u8x16) u16x8_is_equal (x, zero));
+  return u8x16_compare_byte_mask ((u8x16) (x == zero));
 }
 
 always_inline u32
 u32x4_zero_byte_mask (u32x4 x)
 {
   u32x4 zero = { 0 };
-  return u8x16_compare_byte_mask ((u8x16) u32x4_is_equal (x, zero));
+  return u8x16_compare_byte_mask ((u8x16) (x == zero));
 }
 
 always_inline u8x16
diff --git a/src/vppinfra/vhash.h b/src/vppinfra/vhash.h
index 5ab42292001..85dfb788308 100644
--- a/src/vppinfra/vhash.h
+++ b/src/vppinfra/vhash.h
@@ -412,7 +412,7 @@ vhash_bucket_compare (vhash_t * h,
 {
   u32 k = vhash_get_key_word (h, key_word_index, vi);
   u32x4 x = { k, k, k, k };
-  return u32x4_is_equal (bucket[key_word_index].as_u32x4, x);
+  return (bucket[key_word_index].as_u32x4 == x);
 }
 
 #define vhash_bucket_compare_4(h,wi,vi,b0,b1,b2,b3,cmp0,cmp1,cmp2,cmp3)	\
@@ -423,10 +423,10 @@ do {									\
   u32x4 _k2 = u32x4_splat_word (_k4, 2);				\
   u32x4 _k3 = u32x4_splat_word (_k4, 3);				\
 									\
-  cmp0 = u32x4_is_equal (b0->key[wi].as_u32x4, _k0);			\
-  cmp1 = u32x4_is_equal (b1->key[wi].as_u32x4, _k1);			\
-  cmp2 = u32x4_is_equal (b2->key[wi].as_u32x4, _k2);			\
-  cmp3 = u32x4_is_equal (b3->key[wi].as_u32x4, _k3);			\
+  cmp0 = (b0->key[wi].as_u32x4 == _k0);					\
+  cmp1 = (b1->key[wi].as_u32x4 == _k1);					\
+  cmp2 = (b2->key[wi].as_u32x4 == _k2);					\
+  cmp3 = (b3->key[wi].as_u32x4 == _k3);					\
 } while (0)
 
 u32 vhash_get_overflow (vhash_t * h, u32 key_hash, u32 vi, u32 n_key_u32s);