aboutsummaryrefslogtreecommitdiffstats
path: root/src/vnet/classify
diff options
context:
space:
mode:
authorDamjan Marion <damarion@cisco.com>2019-09-25 00:25:36 +0200
committerDave Barach <openvpp@barachs.net>2019-09-26 11:37:30 +0000
commit8304933922620cef005b788a36a4d3f2eab45bb5 (patch)
tree9b873ff3da92bcae785a7a67d7917193ee901b39 /src/vnet/classify
parent6eb009ac023799e4ee1b75d759d726622c39de78 (diff)
classify: use vector code even when data is not aligned
Type: feature Change-Id: I8f5f4841965beb13ebc8c2a37ce0dc331c920109 Signed-off-by: Damjan Marion <damarion@cisco.com>
Diffstat (limited to 'src/vnet/classify')
-rw-r--r--src/vnet/classify/vnet_classify.h268
1 files changed, 126 insertions, 142 deletions
diff --git a/src/vnet/classify/vnet_classify.h b/src/vnet/classify/vnet_classify.h
index 620ef9a1052..35a5db3d27b 100644
--- a/src/vnet/classify/vnet_classify.h
+++ b/src/vnet/classify/vnet_classify.h
@@ -28,8 +28,6 @@ extern vlib_node_registration_t ip6_classify_node;
#define CLASSIFY_TRACE 0
-#define U32X4_ALIGNED(p) PREDICT_TRUE((((intptr_t)p) & 0xf) == 0)
-
/*
* Classify table option to process packets
* CLASSIFY_FLAG_USE_CURR_DATA:
@@ -218,62 +216,57 @@ vnet_classify_hash_packet_inline (vnet_classify_table_t * t, u8 * h)
ASSERT (t);
mask = t->mask;
#ifdef CLIB_HAVE_VEC128
- if (U32X4_ALIGNED (h))
- { //SSE can't handle unaligned data
- u32x4 *data = (u32x4 *) h;
- xor_sum.as_u32x4 = data[0 + t->skip_n_vectors] & mask[0];
- switch (t->match_n_vectors)
- {
- case 5:
- xor_sum.as_u32x4 ^= data[4 + t->skip_n_vectors] & mask[4];
- /* FALLTHROUGH */
- case 4:
- xor_sum.as_u32x4 ^= data[3 + t->skip_n_vectors] & mask[3];
- /* FALLTHROUGH */
- case 3:
- xor_sum.as_u32x4 ^= data[2 + t->skip_n_vectors] & mask[2];
- /* FALLTHROUGH */
- case 2:
- xor_sum.as_u32x4 ^= data[1 + t->skip_n_vectors] & mask[1];
- /* FALLTHROUGH */
- case 1:
- break;
- default:
- abort ();
- }
+ u32x4u *data = (u32x4u *) h;
+ xor_sum.as_u32x4 = data[0 + t->skip_n_vectors] & mask[0];
+ switch (t->match_n_vectors)
+ {
+ case 5:
+ xor_sum.as_u32x4 ^= data[4 + t->skip_n_vectors] & mask[4];
+ /* FALLTHROUGH */
+ case 4:
+ xor_sum.as_u32x4 ^= data[3 + t->skip_n_vectors] & mask[3];
+ /* FALLTHROUGH */
+ case 3:
+ xor_sum.as_u32x4 ^= data[2 + t->skip_n_vectors] & mask[2];
+ /* FALLTHROUGH */
+ case 2:
+ xor_sum.as_u32x4 ^= data[1 + t->skip_n_vectors] & mask[1];
+ /* FALLTHROUGH */
+ case 1:
+ break;
+ default:
+ abort ();
}
- else
-#endif /* CLIB_HAVE_VEC128 */
+#else
+ u32 skip_u64 = t->skip_n_vectors * 2;
+ u64 *data64 = (u64 *) h;
+ xor_sum.as_u64[0] = data64[0 + skip_u64] & ((u64 *) mask)[0];
+ xor_sum.as_u64[1] = data64[1 + skip_u64] & ((u64 *) mask)[1];
+ switch (t->match_n_vectors)
{
- u32 skip_u64 = t->skip_n_vectors * 2;
- u64 *data64 = (u64 *) h;
- xor_sum.as_u64[0] = data64[0 + skip_u64] & ((u64 *) mask)[0];
- xor_sum.as_u64[1] = data64[1 + skip_u64] & ((u64 *) mask)[1];
- switch (t->match_n_vectors)
- {
- case 5:
- xor_sum.as_u64[0] ^= data64[8 + skip_u64] & ((u64 *) mask)[8];
- xor_sum.as_u64[1] ^= data64[9 + skip_u64] & ((u64 *) mask)[9];
- /* FALLTHROUGH */
- case 4:
- xor_sum.as_u64[0] ^= data64[6 + skip_u64] & ((u64 *) mask)[6];
- xor_sum.as_u64[1] ^= data64[7 + skip_u64] & ((u64 *) mask)[7];
- /* FALLTHROUGH */
- case 3:
- xor_sum.as_u64[0] ^= data64[4 + skip_u64] & ((u64 *) mask)[4];
- xor_sum.as_u64[1] ^= data64[5 + skip_u64] & ((u64 *) mask)[5];
- /* FALLTHROUGH */
- case 2:
- xor_sum.as_u64[0] ^= data64[2 + skip_u64] & ((u64 *) mask)[2];
- xor_sum.as_u64[1] ^= data64[3 + skip_u64] & ((u64 *) mask)[3];
- /* FALLTHROUGH */
- case 1:
- break;
-
- default:
- abort ();
- }
+ case 5:
+ xor_sum.as_u64[0] ^= data64[8 + skip_u64] & ((u64 *) mask)[8];
+ xor_sum.as_u64[1] ^= data64[9 + skip_u64] & ((u64 *) mask)[9];
+ /* FALLTHROUGH */
+ case 4:
+ xor_sum.as_u64[0] ^= data64[6 + skip_u64] & ((u64 *) mask)[6];
+ xor_sum.as_u64[1] ^= data64[7 + skip_u64] & ((u64 *) mask)[7];
+ /* FALLTHROUGH */
+ case 3:
+ xor_sum.as_u64[0] ^= data64[4 + skip_u64] & ((u64 *) mask)[4];
+ xor_sum.as_u64[1] ^= data64[5 + skip_u64] & ((u64 *) mask)[5];
+ /* FALLTHROUGH */
+ case 2:
+ xor_sum.as_u64[0] ^= data64[2 + skip_u64] & ((u64 *) mask)[2];
+ xor_sum.as_u64[1] ^= data64[3 + skip_u64] & ((u64 *) mask)[3];
+ /* FALLTHROUGH */
+ case 1:
+ break;
+
+ default:
+ abort ();
}
+#endif /* CLIB_HAVE_VEC128 */
return clib_xxhash (xor_sum.as_u64[0] ^ xor_sum.as_u64[1]);
}
@@ -392,107 +385,98 @@ vnet_classify_find_entry_inline (vnet_classify_table_t * t,
v = vnet_classify_entry_at_index (t, v, value_index);
#ifdef CLIB_HAVE_VEC128
- if (U32X4_ALIGNED (h))
+ u32x4u *data = (u32x4u *) h;
+ for (i = 0; i < limit; i++)
{
- u32x4 *data = (u32x4 *) h;
- for (i = 0; i < limit; i++)
+ key = v->key;
+ result.as_u32x4 = (data[0 + t->skip_n_vectors] & mask[0]) ^ key[0];
+ switch (t->match_n_vectors)
{
- key = v->key;
- result.as_u32x4 = (data[0 + t->skip_n_vectors] & mask[0]) ^ key[0];
- switch (t->match_n_vectors)
- {
- case 5:
- result.as_u32x4 |=
- (data[4 + t->skip_n_vectors] & mask[4]) ^ key[4];
- /* FALLTHROUGH */
- case 4:
- result.as_u32x4 |=
- (data[3 + t->skip_n_vectors] & mask[3]) ^ key[3];
- /* FALLTHROUGH */
- case 3:
- result.as_u32x4 |=
- (data[2 + t->skip_n_vectors] & mask[2]) ^ key[2];
- /* FALLTHROUGH */
- case 2:
- result.as_u32x4 |=
- (data[1 + t->skip_n_vectors] & mask[1]) ^ key[1];
- /* FALLTHROUGH */
- case 1:
- break;
- default:
- abort ();
- }
+ case 5:
+ result.as_u32x4 |= (data[4 + t->skip_n_vectors] & mask[4]) ^ key[4];
+ /* FALLTHROUGH */
+ case 4:
+ result.as_u32x4 |= (data[3 + t->skip_n_vectors] & mask[3]) ^ key[3];
+ /* FALLTHROUGH */
+ case 3:
+ result.as_u32x4 |= (data[2 + t->skip_n_vectors] & mask[2]) ^ key[2];
+ /* FALLTHROUGH */
+ case 2:
+ result.as_u32x4 |= (data[1 + t->skip_n_vectors] & mask[1]) ^ key[1];
+ /* FALLTHROUGH */
+ case 1:
+ break;
+ default:
+ abort ();
+ }
- if (u32x4_zero_byte_mask (result.as_u32x4) == 0xffff)
+ if (u32x4_zero_byte_mask (result.as_u32x4) == 0xffff)
+ {
+ if (PREDICT_TRUE (now))
{
- if (PREDICT_TRUE (now))
- {
- v->hits++;
- v->last_heard = now;
- }
- return (v);
+ v->hits++;
+ v->last_heard = now;
}
- v = vnet_classify_entry_at_index (t, v, 1);
+ return (v);
}
+ v = vnet_classify_entry_at_index (t, v, 1);
}
- else
-#endif /* CLIB_HAVE_VEC128 */
+#else
+ u32 skip_u64 = t->skip_n_vectors * 2;
+ u64 *data64 = (u64 *) h;
+ for (i = 0; i < limit; i++)
{
- u32 skip_u64 = t->skip_n_vectors * 2;
- u64 *data64 = (u64 *) h;
- for (i = 0; i < limit; i++)
- {
- key = v->key;
+ key = v->key;
- result.as_u64[0] =
- (data64[0 + skip_u64] & ((u64 *) mask)[0]) ^ ((u64 *) key)[0];
- result.as_u64[1] =
- (data64[1 + skip_u64] & ((u64 *) mask)[1]) ^ ((u64 *) key)[1];
- switch (t->match_n_vectors)
- {
- case 5:
- result.as_u64[0] |=
- (data64[8 + skip_u64] & ((u64 *) mask)[8]) ^ ((u64 *) key)[8];
- result.as_u64[1] |=
- (data64[9 + skip_u64] & ((u64 *) mask)[9]) ^ ((u64 *) key)[9];
- /* FALLTHROUGH */
- case 4:
- result.as_u64[0] |=
- (data64[6 + skip_u64] & ((u64 *) mask)[6]) ^ ((u64 *) key)[6];
- result.as_u64[1] |=
- (data64[7 + skip_u64] & ((u64 *) mask)[7]) ^ ((u64 *) key)[7];
- /* FALLTHROUGH */
- case 3:
- result.as_u64[0] |=
- (data64[4 + skip_u64] & ((u64 *) mask)[4]) ^ ((u64 *) key)[4];
- result.as_u64[1] |=
- (data64[5 + skip_u64] & ((u64 *) mask)[5]) ^ ((u64 *) key)[5];
- /* FALLTHROUGH */
- case 2:
- result.as_u64[0] |=
- (data64[2 + skip_u64] & ((u64 *) mask)[2]) ^ ((u64 *) key)[2];
- result.as_u64[1] |=
- (data64[3 + skip_u64] & ((u64 *) mask)[3]) ^ ((u64 *) key)[3];
- /* FALLTHROUGH */
- case 1:
- break;
- default:
- abort ();
- }
+ result.as_u64[0] =
+ (data64[0 + skip_u64] & ((u64 *) mask)[0]) ^ ((u64 *) key)[0];
+ result.as_u64[1] =
+ (data64[1 + skip_u64] & ((u64 *) mask)[1]) ^ ((u64 *) key)[1];
+ switch (t->match_n_vectors)
+ {
+ case 5:
+ result.as_u64[0] |=
+ (data64[8 + skip_u64] & ((u64 *) mask)[8]) ^ ((u64 *) key)[8];
+ result.as_u64[1] |=
+ (data64[9 + skip_u64] & ((u64 *) mask)[9]) ^ ((u64 *) key)[9];
+ /* FALLTHROUGH */
+ case 4:
+ result.as_u64[0] |=
+ (data64[6 + skip_u64] & ((u64 *) mask)[6]) ^ ((u64 *) key)[6];
+ result.as_u64[1] |=
+ (data64[7 + skip_u64] & ((u64 *) mask)[7]) ^ ((u64 *) key)[7];
+ /* FALLTHROUGH */
+ case 3:
+ result.as_u64[0] |=
+ (data64[4 + skip_u64] & ((u64 *) mask)[4]) ^ ((u64 *) key)[4];
+ result.as_u64[1] |=
+ (data64[5 + skip_u64] & ((u64 *) mask)[5]) ^ ((u64 *) key)[5];
+ /* FALLTHROUGH */
+ case 2:
+ result.as_u64[0] |=
+ (data64[2 + skip_u64] & ((u64 *) mask)[2]) ^ ((u64 *) key)[2];
+ result.as_u64[1] |=
+ (data64[3 + skip_u64] & ((u64 *) mask)[3]) ^ ((u64 *) key)[3];
+ /* FALLTHROUGH */
+ case 1:
+ break;
+ default:
+ abort ();
+ }
- if (result.as_u64[0] == 0 && result.as_u64[1] == 0)
+ if (result.as_u64[0] == 0 && result.as_u64[1] == 0)
+ {
+ if (PREDICT_TRUE (now))
{
- if (PREDICT_TRUE (now))
- {
- v->hits++;
- v->last_heard = now;
- }
- return (v);
+ v->hits++;
+ v->last_heard = now;
}
-
- v = vnet_classify_entry_at_index (t, v, 1);
+ return (v);
}
+
+ v = vnet_classify_entry_at_index (t, v, 1);
}
+#endif /* CLIB_HAVE_VEC128 */
return 0;
}