The match checks are called inside the search loops, between the transitions; -O3 inlines them and is unchanged, at -O2 they are not. Flatten the search functions so the loops hold no call.
Signed-off-by: Sun Yuechi <[email protected]> --- lib/acl/acl_run_rvv.h | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/lib/acl/acl_run_rvv.h b/lib/acl/acl_run_rvv.h index 1d6fdff045..c4daf10224 100644 --- a/lib/acl/acl_run_rvv.h +++ b/lib/acl/acl_run_rvv.h @@ -182,9 +182,11 @@ transition_vec(vuint32m1_t v_next_input, const uint64_t *trans, } /* - * Execute trie traversal with 8 traversals in parallel + * Execute trie traversal with 8 traversals in parallel. + * Flattened: a vector value live across a call is spilled, and the + * frame is then addressed through vlenb. */ -static inline int +static inline __rte_flatten int search_rvv_8(const struct rte_acl_ctx *ctx, const uint8_t **data, uint32_t *results, uint32_t total_packets, uint32_t categories) { @@ -270,7 +272,7 @@ search_rvv_8(const struct rte_acl_ctx *ctx, const uint8_t **data, /* * Execute trie traversal with 4 traversals in parallel */ -static inline int +static inline __rte_flatten int search_rvv_4(const struct rte_acl_ctx *ctx, const uint8_t **data, uint32_t *results, uint32_t total_packets, uint32_t categories) { -- 2.56.0
