On Wed, Mar 26, 2025 at 11:39:28AM +0100, David Marchand wrote:
> Caught in OBS for Fedora Rawhide on aarch64:
>
> [ 198s] In file included from ../lib/acl/acl_run_neon.h:7,
> [ 198s] from ../lib/acl/acl_run_neon.c:5:
> [ 198s] In function ‘alloc_completion’,
> [ 198s] inlined from ‘acl_start_next_trie’ at
> ../lib/acl/acl_run.h:140:24,
> [ 198s] inlined from ‘search_neon_4.isra’ at
> ../lib/acl/acl_run_neon.h:239:20:
> [ 198s] ../lib/acl/acl_run.h:93:25: error: ‘cmplt’ may be used
> uninitialized [-Werror=maybe-uninitialized]
> [ 198s] 93 | if (p[n].count == 0) {
> [ 198s] | ~~~~^~~~~~
> [ 198s] ../lib/acl/acl_run_neon.h: In function ‘search_neon_4.isra’:
> [ 198s] ../lib/acl/acl_run_neon.h:230:27: note: ‘cmplt’ declared here
> [ 198s] 230 | struct completion cmplt[4];
> [ 198s] | ^~~~~
>
> The code was resetting sequentially cmpl[].count at the exact index that
> later call to alloc_completion uses.
> While this code seems correct, GCC 15 does not understand this (probably
> when applying some optimisations).
>
> Instead, reset cmpl[].count all at once in acl_set_flow, and cleanup the
> various vectorized implementations accordingly.
>
> Bugzilla ID: 1678
> Cc: sta...@dpdk.org
>
> Signed-off-by: David Marchand <david.march...@redhat.com>
> ---
> lib/acl/acl_run.h | 5 +++++
> lib/acl/acl_run_altivec.h | 8 ++------
> lib/acl/acl_run_avx2.h | 4 +---
> lib/acl/acl_run_neon.h | 8 ++------
> lib/acl/acl_run_scalar.c | 4 +---
> lib/acl/acl_run_sse.h | 8 ++------
> 6 files changed, 13 insertions(+), 24 deletions(-)
>
> diff --git a/lib/acl/acl_run.h b/lib/acl/acl_run.h
> index 7f092413cd..9fd3e60021 100644
> --- a/lib/acl/acl_run.h
> +++ b/lib/acl/acl_run.h
> @@ -176,6 +176,8 @@ acl_set_flow(struct acl_flow_data *flows, struct
> completion *cmplt,
> uint32_t cmplt_size, const uint8_t **data, uint32_t *results,
> uint32_t data_num, uint32_t categories, const uint64_t *trans)
> {
> + unsigned int i;
> +
> flows->num_packets = 0;
> flows->started = 0;
> flows->trie = 0;
> @@ -187,6 +189,9 @@ acl_set_flow(struct acl_flow_data *flows, struct
> completion *cmplt,
> flows->data = data;
> flows->results = results;
> flows->trans = trans;
> +
> + for (i = 0; i < cmplt_size; i++)
> + cmplt[i].count = 0;
> }
Minor nit, but since we are using c11 standard, is it not better to declare
"i" inside the "for" statement. Keeps diffs simpler for adding/removing
code, I think.
For the rest of the code, I need to take a bit more time reviewing to be
sure I understand the change. I'll try and get to it later.
/Bruce
>
> typedef void (*resolve_priority_t)
> diff --git a/lib/acl/acl_run_altivec.h b/lib/acl/acl_run_altivec.h
> index 2d398ffded..d5ccdb94f0 100644
> --- a/lib/acl/acl_run_altivec.h
> +++ b/lib/acl/acl_run_altivec.h
> @@ -199,10 +199,8 @@ search_altivec_8(const struct rte_acl_ctx *ctx, const
> uint8_t **data,
> acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
> total_packets, categories, ctx->trans_table);
>
> - for (n = 0; n < MAX_SEARCHES_ALTIVEC8; n++) {
> - cmplt[n].count = 0;
> + for (n = 0; n < MAX_SEARCHES_ALTIVEC8; n++)
> index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
> - }
>
<snip>
