From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id EA55146DD5; Tue, 26 Aug 2025 08:55:25 +0200 (CEST) Received: from mails.dpdk.org (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 989244021E; Tue, 26 Aug 2025 08:55:25 +0200 (CEST) Received: from dkmailrelay1.smartsharesystems.com (smartserver.smartsharesystems.com [77.243.40.215]) by mails.dpdk.org (Postfix) with ESMTP id 8EB6840156 for ; Tue, 26 Aug 2025 08:55:24 +0200 (CEST) Received: from smartserver.smartsharesystems.com (smartserver.smartsharesys.local [192.168.4.10]) by dkmailrelay1.smartsharesystems.com (Postfix) with ESMTP id 3C58D2071D; Tue, 26 Aug 2025 08:55:24 +0200 (CEST) Subject: RE: [PATCH v2 3/4] hash: reduce architecture special cases MIME-Version: 1.0 Content-Type: text/plain; charset="iso-8859-1" Content-Transfer-Encoding: quoted-printable Date: Tue, 26 Aug 2025 08:55:23 +0200 Message-ID: <98CBD80474FA8B44BF855DF32C47DC35E9FE6C@smartserver.smartshare.dk> Content-class: urn:content-classes:message In-Reply-To: <20250822182110.27599-4-stephen@networkplumber.org> X-MS-Has-Attach: X-MS-TNEF-Correlator: Thread-Topic: [PATCH v2 3/4] hash: reduce architecture special cases X-MimeOLE: Produced By Microsoft Exchange V6.5 Thread-Index: AdwTkZq5RrPTkRzrSkSB8nruUgp0EACv6+pw References: <20250821203646.133506-1-stephen@networkplumber.org> <20250822182110.27599-1-stephen@networkplumber.org> <20250822182110.27599-4-stephen@networkplumber.org> From: =?iso-8859-1?Q?Morten_Br=F8rup?= To: "Stephen Hemminger" , Cc: "Wathsala Vithanage" , "Yipeng Wang" , "Sameh Gobriel" , "Bruce Richardson" , "Vladimir Medvedkin" , =?iso-8859-1?Q?Mattias_R=F6nnblom?= X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org > From: Stephen Hemminger [mailto:stephen@networkplumber.org] > Sent: Friday, 22 August 2025 20.20 >=20 > Make comparison of sizes compatible across platforms. > Keep the special case code for 16 bytes for x86 and arm64 but > also add simple xor for others. >=20 > Need to keep rte_hash_k32_cmp_eq() exposed because ip_frag > code poaches it. >=20 > Signed-off-by: Stephen Hemminger > --- > lib/hash/rte_cmp_arm64.h | 56 +------------------------ > lib/hash/rte_cmp_generic.h | 35 ++++++++++++++++ > lib/hash/rte_cmp_x86.h | 60 ++------------------------ > lib/hash/rte_cuckoo_hash.c | 86 = +++++++++++++++++++++++++++++++++----- > 4 files changed, 116 insertions(+), 121 deletions(-) > create mode 100644 lib/hash/rte_cmp_generic.h >=20 > diff --git a/lib/hash/rte_cmp_arm64.h b/lib/hash/rte_cmp_arm64.h > index a3e85635eb..2b2a37ebd2 100644 > --- a/lib/hash/rte_cmp_arm64.h > +++ b/lib/hash/rte_cmp_arm64.h > @@ -2,7 +2,7 @@ > * Copyright(c) 2015 Cavium, Inc > */ >=20 > -/* Functions to compare multiple of 16 byte keys (up to 128 bytes) */ > +/* Functions to compare multiple of 16 byte keys */ > static inline int > rte_hash_k16_cmp_eq(const void *key1, const void *key2, > size_t key_len __rte_unused) > @@ -27,59 +27,7 @@ rte_hash_k16_cmp_eq(const void *key1, const void > *key2, > static inline int > rte_hash_k32_cmp_eq(const void *key1, const void *key2, size_t = key_len) > { > - return rte_hash_k16_cmp_eq(key1, key2, key_len) || > + return rte_hash_k16_cmp_eq(key1, key2, key_len) | > rte_hash_k16_cmp_eq((const char *) key1 + 16, > (const char *) key2 + 16, key_len); > } Agree with the change from "||" to "|", to compare in blocks of 32 bytes = instead of 16. Another potential improvement: Consider passing the actual length, instead of passing key_len, i.e.: return rte_hash_k16_cmp_eq(key1, key2, 16) | rte_hash_k16_cmp_eq((const char *) key1 + 16, (const char *) key2 + 16, 16); [...] > diff --git a/lib/hash/rte_cmp_x86.h b/lib/hash/rte_cmp_x86.h > index ddfbef462f..e7a38c8fcd 100644 > --- a/lib/hash/rte_cmp_x86.h > +++ b/lib/hash/rte_cmp_x86.h > @@ -4,7 +4,7 @@ >=20 > #include >=20 > -/* Functions to compare multiple of 16 byte keys (up to 128 bytes) */ > +/* Function to compare multiple of 16 byte keys */ > static inline int > rte_hash_k16_cmp_eq(const void *key1, const void *key2, size_t = key_len > __rte_unused) > { > @@ -18,59 +18,7 @@ rte_hash_k16_cmp_eq(const void *key1, const void > *key2, size_t key_len __rte_unu > static inline int > rte_hash_k32_cmp_eq(const void *key1, const void *key2, size_t = key_len) > { > - return rte_hash_k16_cmp_eq(key1, key2, key_len) || > - rte_hash_k16_cmp_eq((const char *) key1 + 16, > - (const char *) key2 + 16, key_len); > -} > - > -static inline int > -rte_hash_k48_cmp_eq(const void *key1, const void *key2, size_t = key_len) > -{ > - return rte_hash_k16_cmp_eq(key1, key2, key_len) || > - rte_hash_k16_cmp_eq((const char *) key1 + 16, > - (const char *) key2 + 16, key_len) || > - rte_hash_k16_cmp_eq((const char *) key1 + 32, > - (const char *) key2 + 32, key_len); > -} > - > -static inline int > -rte_hash_k64_cmp_eq(const void *key1, const void *key2, size_t = key_len) > -{ > - return rte_hash_k32_cmp_eq(key1, key2, key_len) || > - rte_hash_k32_cmp_eq((const char *) key1 + 32, > - (const char *) key2 + 32, key_len); > -} > - > -static inline int > -rte_hash_k80_cmp_eq(const void *key1, const void *key2, size_t = key_len) > -{ > - return rte_hash_k64_cmp_eq(key1, key2, key_len) || > - rte_hash_k16_cmp_eq((const char *) key1 + 64, > - (const char *) key2 + 64, key_len); > -} > - > -static inline int > -rte_hash_k96_cmp_eq(const void *key1, const void *key2, size_t = key_len) > -{ > - return rte_hash_k64_cmp_eq(key1, key2, key_len) || > - rte_hash_k32_cmp_eq((const char *) key1 + 64, > - (const char *) key2 + 64, key_len); > -} > - > -static inline int > -rte_hash_k112_cmp_eq(const void *key1, const void *key2, size_t > key_len) > -{ > - return rte_hash_k64_cmp_eq(key1, key2, key_len) || > - rte_hash_k32_cmp_eq((const char *) key1 + 64, > - (const char *) key2 + 64, key_len) || > - rte_hash_k16_cmp_eq((const char *) key1 + 96, > - (const char *) key2 + 96, key_len); > -} > - > -static inline int > -rte_hash_k128_cmp_eq(const void *key1, const void *key2, size_t > key_len) > -{ > - return rte_hash_k64_cmp_eq(key1, key2, key_len) || > - rte_hash_k64_cmp_eq((const char *) key1 + 64, > - (const char *) key2 + 64, key_len); > + return rte_hash_k16_cmp_eq(key1, key2, key_len) | > + rte_hash_k16_cmp_eq((const uint8_t *) key1 + 16, > + (const uint8_t *) key2 + 16, key_len); You should use "const char *" (not "const uint8_t *") here too. Or add RTE_CONST_PTR_ADD() to rte_common.h: /** * add a byte-value offset to a const pointer */ #define RTE_CONST_PTR_ADD(ptr, x) ((const void*)((uintptr_t)(ptr) + = (x))) > } > diff --git a/lib/hash/rte_cuckoo_hash.c b/lib/hash/rte_cuckoo_hash.c > index 619fe0c691..199cb62bf0 100644 > --- a/lib/hash/rte_cuckoo_hash.c > +++ b/lib/hash/rte_cuckoo_hash.c > @@ -42,13 +42,6 @@ RTE_LOG_REGISTER_DEFAULT(hash_logtype, INFO); > #define RETURN_IF_TRUE(cond, retval) > #endif >=20 > -#if defined(RTE_ARCH_X86) > -#include "rte_cmp_x86.h" > -#endif > - > -#if defined(RTE_ARCH_ARM64) > -#include "rte_cmp_arm64.h" > -#endif >=20 > /* > * All different options to select a key compare function, > @@ -57,7 +50,6 @@ RTE_LOG_REGISTER_DEFAULT(hash_logtype, INFO); > */ > enum cmp_jump_table_case { > KEY_CUSTOM =3D 0, > -#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64) > KEY_16_BYTES, > KEY_32_BYTES, > KEY_48_BYTES, > @@ -66,11 +58,85 @@ enum cmp_jump_table_case { > KEY_96_BYTES, > KEY_112_BYTES, > KEY_128_BYTES, > -#endif > KEY_OTHER_BYTES, > NUM_KEY_CMP_CASES, > }; >=20 > +/* > + * Comparison functions for different key sizes. > + * Each function is only called with a specific fixed key size. > + * > + * Return value is 0 on equality to allow direct use of memcmp. > + * Recommend using XOR and | operator to avoid branching > + * as long as key is smaller than cache line size. > + * > + * Key1 always points to key[] in rte_hash_key which is aligned. > + * Key2 is parameter to insert which might not be. > + * > + * Special case for 16 and 32 bytes to allow for architecture > + * specific optimizations. > + */ > + > +#if defined(RTE_ARCH_X86) > +#include "rte_cmp_x86.h" > +#elif defined(RTE_ARCH_ARM64) > +#include "rte_cmp_arm64.h" > +#else > +#include "rte_cmp_generic.h" > +#endif > + > +static int > +rte_hash_k48_cmp_eq(const void *key1, const void *key2, size_t = key_len) > +{ > + return rte_hash_k16_cmp_eq(key1, key2, key_len) | > + rte_hash_k16_cmp_eq((const uint8_t *) key1 + 16, > + (const uint8_t *) key2 + 16, key_len) || > + rte_hash_k16_cmp_eq((const uint8_t *) key1 + 32, > + (const uint8_t *) key2 + 32, key_len); > +} > + > +static int > +rte_hash_k64_cmp_eq(const void *key1, const void *key2, size_t = key_len) > +{ > + return rte_hash_k32_cmp_eq(key1, key2, key_len) | Is the "|" instead of "||", to compare in blocks of 64 bytes instead of = 32, intentional? > + rte_hash_k32_cmp_eq((const uint8_t *) key1 + 32, > + (const uint8_t *) key2 + 32, key_len); > +} > + > +static int > +rte_hash_k80_cmp_eq(const void *key1, const void *key2, size_t = key_len) > +{ > + return rte_hash_k64_cmp_eq(key1, key2, key_len) || > + rte_hash_k16_cmp_eq((const uint8_t *) key1 + 64, > + (const uint8_t *) key2 + 64, key_len); > +} > + > +static int > +rte_hash_k96_cmp_eq(const void *key1, const void *key2, size_t = key_len) > +{ > + return rte_hash_k64_cmp_eq(key1, key2, key_len) || > + rte_hash_k32_cmp_eq((const uint8_t *) key1 + 64, > + (const uint8_t *) key2 + 64, key_len); > +} > + > +static int > +rte_hash_k112_cmp_eq(const void *key1, const void *key2, size_t > key_len) > +{ > + return rte_hash_k64_cmp_eq(key1, key2, key_len) || > + rte_hash_k32_cmp_eq((const uint8_t *) key1 + 64, > + (const uint8_t *) key2 + 64, key_len) || > + rte_hash_k16_cmp_eq((const uint8_t *) key1 + 96, > + (const uint8_t *) key2 + 96, key_len); > +} > + > +static int > +rte_hash_k128_cmp_eq(const void *key1, const void *key2, size_t > key_len) > +{ > + return rte_hash_k64_cmp_eq(key1, key2, key_len) || > + rte_hash_k64_cmp_eq((const uint8_t *) key1 + 64, > + (const uint8_t *) key2 + 64, key_len); > +} > + > /* Enum used to select the implementation of the signature comparison > function to use > * eg: a system supporting SVE might want to use a NEON or scalar > implementation. > */ > @@ -160,7 +226,6 @@ void rte_hash_set_cmp_func(struct rte_hash *h, > rte_hash_cmp_eq_t func) > */ > static const rte_hash_cmp_eq_t cmp_jump_table[NUM_KEY_CMP_CASES] =3D = { > [KEY_CUSTOM] =3D NULL, > -#if defined(RTE_ARCH_X86) || defined(RTE_ARCH_ARM64) > [KEY_16_BYTES] =3D rte_hash_k16_cmp_eq, > [KEY_32_BYTES] =3D rte_hash_k32_cmp_eq, > [KEY_48_BYTES] =3D rte_hash_k48_cmp_eq, > @@ -169,7 +234,6 @@ static const rte_hash_cmp_eq_t > cmp_jump_table[NUM_KEY_CMP_CASES] =3D { > [KEY_96_BYTES] =3D rte_hash_k96_cmp_eq, > [KEY_112_BYTES] =3D rte_hash_k112_cmp_eq, > [KEY_128_BYTES] =3D rte_hash_k128_cmp_eq, > -#endif > [KEY_OTHER_BYTES] =3D memcmp, > }; >=20 > -- > 2.47.2 With or without suggested changes... Acked-by: Morten Br=F8rup