The current implementation attempts to use a uint16_t to alias the lpm table
structures. Such aliasing can break optimizer performance. This patch uses
union type indirection and adds static inline functions for performing the
aliasing.
Signed-off-by: Aaron Conole <aconole at redhat.com>
---
lib/librte_lpm/rte_lpm.h | 53 +++++++++++++++++++++++++++++++++++++-----------
1 file changed, 41 insertions(+), 12 deletions(-)
diff --git a/lib/librte_lpm/rte_lpm.h b/lib/librte_lpm/rte_lpm.h
index c299ce2..eae6ff1 100644
--- a/lib/librte_lpm/rte_lpm.h
+++ b/lib/librte_lpm/rte_lpm.h
@@ -157,6 +157,33 @@ struct rte_lpm {
};
/**
+ * Convert from tbl_entry types to integer types
+ */
+static inline uint16_t
+rte_lpm_tbl24_entry_to_uint16(const struct rte_lpm_tbl24_entry *entry)
+{
+ union {
+ uint16_t i;
+ struct rte_lpm_tbl24_entry s;
+ } tbl_entry_u;
+
+ tbl_entry_u.s = *entry;
+ return tbl_entry_u.i;
+}
+
+static inline uint16_t
+rte_lpm_tbl8_entry_to_uint16(const struct rte_lpm_tbl8_entry *entry)
+{
+ union {
+ uint16_t i;
+ struct rte_lpm_tbl8_entry s;
+ } tbl_entry_u;
+
+ tbl_entry_u.s = *entry;
+ return tbl_entry_u.i;
+}
+
+/**
* Create an LPM object.
*
* @param name
@@ -286,7 +313,7 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip, uint8_t
*next_hop)
RTE_LPM_RETURN_IF_TRUE(((lpm == NULL) || (next_hop == NULL)), -EINVAL);
/* Copy tbl24 entry */
- tbl_entry = *(const uint16_t *)&lpm->tbl24[tbl24_index];
+ tbl_entry = rte_lpm_tbl24_entry_to_uint16(&lpm->tbl24[tbl24_index]);
/* Copy tbl8 entry (only if needed) */
if (unlikely((tbl_entry & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
@@ -295,7 +322,7 @@ rte_lpm_lookup(struct rte_lpm *lpm, uint32_t ip, uint8_t
*next_hop)
unsigned tbl8_index = (uint8_t)ip +
((uint8_t)tbl_entry *
RTE_LPM_TBL8_GROUP_NUM_ENTRIES);
- tbl_entry = *(const uint16_t *)&lpm->tbl8[tbl8_index];
+ tbl_entry =
rte_lpm_tbl8_entry_to_uint16(&lpm->tbl8[tbl8_index]);
}
*next_hop = (uint8_t)tbl_entry;
@@ -342,7 +369,8 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const
uint32_t * ips,
for (i = 0; i < n; i++) {
/* Simply copy tbl24 entry to output */
- next_hops[i] = *(const uint16_t *)&lpm->tbl24[tbl24_indexes[i]];
+ next_hops[i] = rte_lpm_tbl24_entry_to_uint16(
+ &lpm->tbl24[tbl24_indexes[i]]);
/* Overwrite output with tbl8 entry if needed */
if (unlikely((next_hops[i] & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
@@ -352,7 +380,8 @@ rte_lpm_lookup_bulk_func(const struct rte_lpm *lpm, const
uint32_t * ips,
((uint8_t)next_hops[i] *
RTE_LPM_TBL8_GROUP_NUM_ENTRIES);
- next_hops[i] = *(const uint16_t
*)&lpm->tbl8[tbl8_index];
+ next_hops[i] = rte_lpm_tbl8_entry_to_uint16(
+ &lpm->tbl8[tbl8_index]);
}
}
return 0;
@@ -419,13 +448,13 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip,
uint16_t hop[4],
idx = _mm_cvtsi128_si64(i24);
i24 = _mm_srli_si128(i24, sizeof(uint64_t));
- tbl[0] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx];
- tbl[1] = *(const uint16_t *)&lpm->tbl24[idx >> 32];
+ tbl[0] = rte_lpm_tbl24_entry_to_uint16(&lpm->tbl24[(uint32_t)idx]);
+ tbl[1] = rte_lpm_tbl24_entry_to_uint16(&lpm->tbl24[idx >> 32]);
idx = _mm_cvtsi128_si64(i24);
- tbl[2] = *(const uint16_t *)&lpm->tbl24[(uint32_t)idx];
- tbl[3] = *(const uint16_t *)&lpm->tbl24[idx >> 32];
+ tbl[2] = rte_lpm_tbl24_entry_to_uint16(&lpm->tbl24[(uint32_t)idx]);
+ tbl[3] = rte_lpm_tbl24_entry_to_uint16(&lpm->tbl24[idx >> 32]);
/* get 4 indexes for tbl8[]. */
i8.x = _mm_and_si128(ip, mask8);
@@ -446,25 +475,25 @@ rte_lpm_lookupx4(const struct rte_lpm *lpm, __m128i ip,
uint16_t hop[4],
RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
i8.u32[0] = i8.u32[0] +
(uint8_t)tbl[0] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
- tbl[0] = *(const uint16_t *)&lpm->tbl8[i8.u32[0]];
+ tbl[0] = rte_lpm_tbl8_entry_to_uint16(&lpm->tbl8[i8.u32[0]]);
}
if (unlikely((pt >> 16 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
i8.u32[1] = i8.u32[1] +
(uint8_t)tbl[1] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
- tbl[1] = *(const uint16_t *)&lpm->tbl8[i8.u32[1]];
+ tbl[1] = rte_lpm_tbl8_entry_to_uint16(&lpm->tbl8[i8.u32[1]]);
}
if (unlikely((pt >> 32 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
i8.u32[2] = i8.u32[2] +
(uint8_t)tbl[2] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
- tbl[2] = *(const uint16_t *)&lpm->tbl8[i8.u32[2]];
+ tbl[2] = rte_lpm_tbl8_entry_to_uint16(&lpm->tbl8[i8.u32[2]]);
}
if (unlikely((pt >> 48 & RTE_LPM_VALID_EXT_ENTRY_BITMASK) ==
RTE_LPM_VALID_EXT_ENTRY_BITMASK)) {
i8.u32[3] = i8.u32[3] +
(uint8_t)tbl[3] * RTE_LPM_TBL8_GROUP_NUM_ENTRIES;
- tbl[3] = *(const uint16_t *)&lpm->tbl8[i8.u32[3]];
+ tbl[3] = rte_lpm_tbl8_entry_to_uint16(&lpm->tbl8[i8.u32[3]]);
}
hop[0] = (tbl[0] & RTE_LPM_LOOKUP_SUCCESS) ? (uint8_t)tbl[0] : defv;
--
2.5.0
