On Wed, Nov 27, 2024 at 02:57:57PM -0800, Andre Muezerie wrote:
> The mechanism used to initialize an __m128i data type in rte_thash.h is
> non-portable and MSVC does not like it. It clearly is not doing what
> is desired:
>
> ..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
> truncation from 'unsigned __int64' to 'char'
> ..\lib\hash\rte_thash.h(38): warning C4305: 'initializing':
> truncation from 'unsigned __int64' to 'char'
>
> A more portable approach is to use compiler intrinsics to perform the
> initialization. This patch uses a single compiler intrinsic to
> initialize the data type using a sequence of 16 bytes stored in
> memory.
>
> There should be no perf degradation due to this change.
>
> Signed-off-by: Andre Muezerie <andre...@linux.microsoft.com>
> ---
> lib/hash/rte_thash.h | 11 +++++++----
> 1 file changed, 7 insertions(+), 4 deletions(-)
>
> diff --git a/lib/hash/rte_thash.h b/lib/hash/rte_thash.h
> index c0af5968df..3512639792 100644
> --- a/lib/hash/rte_thash.h
> +++ b/lib/hash/rte_thash.h
> @@ -34,8 +34,9 @@ extern "C" {
> /* Byte swap mask used for converting IPv6 address
> * 4-byte chunks to CPU byte order
> */
> -static const __m128i rte_thash_ipv6_bswap_mask = {
> - 0x0405060700010203ULL, 0x0C0D0E0F08090A0BULL};
> +static const uint8_t rte_thash_ipv6_bswap_mask[] = {
> + 0x03, 0x02, 0x01, 0x00, 0x07, 0x06, 0x05, 0x04,
> + 0x0B, 0x0A, 0x09, 0x08, 0x0F, 0x0E, 0x0D, 0x0C};
> #endif
>
> /**
> @@ -152,12 +153,14 @@ rte_thash_load_v6_addrs(const struct rte_ipv6_hdr *orig,
> union rte_thash_tuple *targ)
> {
> #ifdef RTE_ARCH_X86
> + const __m128i ipv6_bswap_mask =
> + _mm_loadu_si128((const
> __m128i*)&rte_thash_ipv6_bswap_mask);
> __m128i ipv6 = _mm_loadu_si128((const __m128i *)&orig->src_addr);
> *(__m128i *)&targ->v6.src_addr =
> - _mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
> + _mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
> ipv6 = _mm_loadu_si128((const __m128i *)&orig->dst_addr);
> *(__m128i *)&targ->v6.dst_addr =
> - _mm_shuffle_epi8(ipv6, rte_thash_ipv6_bswap_mask);
> + _mm_shuffle_epi8(ipv6, ipv6_bswap_mask);
> #elif defined(__ARM_NEON)
> uint8x16_t ipv6 = vld1q_u8(orig->src_addr.a);
> vst1q_u8(targ->v6.src_addr.a, vrev32q_u8(ipv6));
> --
> 2.34.1
Could someone please review this patch and let me know if there are changes to
be made?
I have other patches depending on this.
Thanks,
Andre Muezerie
