vppinfra: add 128-bit and 512-bit a ^ b ^ c shortcut

This allows us to combine 2 XOR operations into signle instruction
which makes difference in crypto op:

- in x86, by using ternary logic instruction
- on ARM, by using EOR3 instruction (available with sha3 feature)

Type: refactor
Change-Id: Ibdf9001840399d2f838d491ca81b57cbd8430433
Signed-off-by: Damjan Marion <damjan.marion@gmail.com>

4 files changed, 33 insertions, 18 deletions

diff --git a/src/plugins/crypto_native/ghash.h b/src/plugins/crypto_native/ghash.h
index 1ee1a997997..a2886a468e9 100644
--- a/src/plugins/crypto_native/ghash.h
+++ b/src/plugins/crypto_native/ghash.h
@@ -105,18 +105,6 @@
 #ifndef __ghash_h__
 #define __ghash_h__
 
-/* on AVX-512 systems we can save a clock cycle by using ternary logic
-   instruction to calculate a XOR b XOR c */
-static_always_inline u8x16
-ghash_xor3 (u8x16 a, u8x16 b, u8x16 c)
-{
-#if defined (__AVX512F__)
-  return (u8x16) _mm_ternarylogic_epi32 ((__m128i) a, (__m128i) b,
-					 (__m128i) c, 0x96);
-#endif
-  return a ^ b ^ c;
-}
-
 static_always_inline u8x16
 gmul_lo_lo (u8x16 a, u8x16 b)
 {
@@ -204,8 +192,8 @@ ghash_mul_next (ghash_data_t * gd, u8x16 a, u8x16 b)
   if (gd->pending)
     {
       /* there is peding data from previous invocation so we can XOR */
-      gd->hi = ghash_xor3 (gd->hi, gd->tmp_hi, hi);
-      gd->lo = ghash_xor3 (gd->lo, gd->tmp_lo, lo);
+      gd->hi = u8x16_xor3 (gd->hi, gd->tmp_hi, hi);
+      gd->lo = u8x16_xor3 (gd->lo, gd->tmp_lo, lo);
       gd->pending = 0;
     }
   else
@@ -217,7 +205,7 @@ ghash_mul_next (ghash_data_t * gd, u8x16 a, u8x16 b)
     }
 
   /* gd->mid ^= a0 * b1 ^ a1 * b0  */
-  gd->mid = ghash_xor3 (gd->mid, gmul_hi_lo (a, b), gmul_lo_hi (a, b));
+  gd->mid = u8x16_xor3 (gd->mid, gmul_hi_lo (a, b), gmul_lo_hi (a, b));
 }
 
 static_always_inline void
@@ -233,8 +221,8 @@ ghash_reduce (ghash_data_t * gd)
 
   if (gd->pending)
     {
-      gd->lo = ghash_xor3 (gd->lo, gd->tmp_lo, midl);
-      gd->hi = ghash_xor3 (gd->hi, gd->tmp_hi, midr);
+      gd->lo = u8x16_xor3 (gd->lo, gd->tmp_lo, midl);
+      gd->hi = u8x16_xor3 (gd->hi, gd->tmp_hi, midr);
     }
   else
     {
@@ -255,7 +243,7 @@ ghash_reduce2 (ghash_data_t * gd)
 static_always_inline u8x16
 ghash_final (ghash_data_t * gd)
 {
-  return ghash_xor3 (gd->hi, u8x16_word_shift_right (gd->tmp_lo, 4),
+  return u8x16_xor3 (gd->hi, u8x16_word_shift_right (gd->tmp_lo, 4),
 		     u8x16_word_shift_left (gd->tmp_hi, 4));
 }
 
diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h
index c54d8cd2499..29d96f85ce9 100644
--- a/src/vppinfra/vector_avx512.h
+++ b/src/vppinfra/vector_avx512.h
@@ -143,6 +143,13 @@ u64x8_permute (u64x8 a, u64x8 b, u64x8 mask)
 #define u32x16_ternary_logic(a, b, c, d) \
   (u32x16) _mm512_ternarylogic_epi32 ((__m512i) a, (__m512i) b, (__m512i) c, d)
 
+static_always_inline u8x64
+u8x64_xor3 (u8x64 a, u8x64 b, u8x64 c)
+{
+  return (u8x64) _mm512_ternarylogic_epi32 ((__m512i) a, (__m512i) b,
+					    (__m512i) c, 0x96);
+}
+
 static_always_inline void
 u32x16_transpose (u32x16 m[16])
 {
diff --git a/src/vppinfra/vector_neon.h b/src/vppinfra/vector_neon.h
index 81d99a64f05..3855f55ad41 100644
--- a/src/vppinfra/vector_neon.h
+++ b/src/vppinfra/vector_neon.h
@@ -203,6 +203,17 @@ u8x16_reflect (u8x16 v)
   return (u8x16) vqtbl1q_u8 (v, mask);
 }
 
+static_always_inline u8x16
+u8x16_xor3 (u8x16 a, u8x16 b, u8x16 c)
+{
+#if __GNUC__ == 8 && __ARM_FEATURE_SHA3 == 1
+  u8x16 r;
+__asm__ ("eor3 %0.16b,%1.16b,%2.16b,%3.16b": "=w" (r): "0" (a), "w" (b), "w" (c):);
+  return r;
+#endif
+  return a ^ b ^ c;
+}
+
 #define CLIB_HAVE_VEC128_MSB_MASK
 
 #define CLIB_HAVE_VEC128_UNALIGNED_LOAD_STORE
diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h
index c22e86e7437..e75580e6026 100644
--- a/src/vppinfra/vector_sse42.h
+++ b/src/vppinfra/vector_sse42.h
@@ -746,6 +746,15 @@ u8x16_blend (u8x16 v1, u8x16 v2, u8x16 mask)
   return (u8x16) _mm_blendv_epi8 ((__m128i) v1, (__m128i) v2, (__m128i) mask);
 }
 
+static_always_inline u8x16
+u8x16_xor3 (u8x16 a, u8x16 b, u8x16 c)
+{
+#if __AVX512F__
+  return (u8x16) _mm_ternarylogic_epi32 ((__m128i) a, (__m128i) b,
+					 (__m128i) c, 0x96);
+#endif
+  return a ^ b ^ c;
+}
 
 #endif /* included_vector_sse2_h */