From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pl1-x62a.google.com (mail-pl1-x62a.google.com [IPv6:2607:f8b0:4864:20::62a]) by sourceware.org (Postfix) with ESMTPS id 3D962385741A for ; Thu, 24 Mar 2022 18:56:52 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org 3D962385741A Received: by mail-pl1-x62a.google.com with SMTP id n18so5712114plg.5 for ; Thu, 24 Mar 2022 11:56:52 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20210112; h=x-gm-message-state:mime-version:references:in-reply-to:from:date :message-id:subject:to:cc; bh=Cdo4noTL0wKRiKyxAJIu2BkeTzLn6wqBwwEJxPaBeeU=; b=Ejxd3t+jIKjWRojR09R55F/HB5zkkhvq96AZLgALYiiQvGWhpoZ3j1wSBXXO7fkx1V h2xDwJjOz+L0knZBHSE658KWrTaKMe9ZEiWULWow2rf3iyVOG94OQNK/6Pqd9OIPUp5Z qCb8KHwYO/s/k1k88vFtJxg6RgK53BKQkm7UlXMMjNizR5k0HnBs27PJc305L3U0Y3iL GWLyEVwdytiBHuOIuFVJvRBTMCmRQXyISfJAMD0VxuNtT9vyjYuTUL9tawe+ZU6/wHF7 2TEQrZ1jGH/eM4d47mJ+0EB+aTyaqkIT3lab3h/DbMEHZI22uf9K7s4Mf1ocy7xUDHJ2 XKJQ== X-Gm-Message-State: AOAM530C9JrkWwUH38j+s8jr4jkKrkisNagJVjccg9tuhRDSOxWWw99v 41Ej0oGM+XQt9Cklu0Ob/GufU/zQMkN2KlDzCy0= X-Google-Smtp-Source: ABdhPJx7Thmiqx6yJTKLclyuq4AtO6/m/LxtQMlKlcWLlchBtzTEn5vxaQAIoL4LUnbE4IqiR5ljZ2Lv2c7K/jxjoY0= X-Received: by 2002:a17:902:f690:b0:154:2ceb:f18a with SMTP id l16-20020a170902f69000b001542cebf18amr7421318plg.4.1648148210991; Thu, 24 Mar 2022 11:56:50 -0700 (PDT) MIME-Version: 1.0 References: <20220323215734.3927131-1-goldstein.w.n@gmail.com> <20220323215734.3927131-8-goldstein.w.n@gmail.com> In-Reply-To: <20220323215734.3927131-8-goldstein.w.n@gmail.com> From: "H.J. Lu" Date: Thu, 24 Mar 2022 11:56:15 -0700 Message-ID: Subject: Re: [PATCH v1 08/23] x86: Optimize strspn in strspn-c.c To: Noah Goldstein Cc: GNU C Library , "Carlos O'Donell" Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3026.4 required=5.0 tests=BAYES_00, DKIM_SIGNED, DKIM_VALID, DKIM_VALID_AU, DKIM_VALID_EF, FREEMAIL_FROM, GIT_PATCH_0, RCVD_IN_DNSWL_NONE, SPF_HELO_NONE, SPF_PASS, TXREP, T_SCC_BODY_TEXT_LINE autolearn=ham autolearn_force=no version=3.4.4 X-Spam-Checker-Version: SpamAssassin 3.4.4 (2020-01-24) on server2.sourceware.org X-BeenThere: libc-alpha@sourceware.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: Libc-alpha mailing list List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , X-List-Received-Date: Thu, 24 Mar 2022 18:56:55 -0000 On Wed, Mar 23, 2022 at 2:59 PM Noah Goldstein wrote: > > Use _mm_cmpeq_epi8 and _mm_movemask_epi8 to get strlen instead of > _mm_cmpistri. Also change offset to unsigned to avoid unnecessary > sign extensions. > > geometric_mean(N=20) of all benchmarks that dont fallback on > sse2; New / Original: .901 > > All string/memory tests pass. > --- > Geomtric Mean N=20 runs; All functions page aligned > len, align1, align2, pos, New Time / Old Time > 1, 0, 0, 512, 0.768 > 1, 1, 0, 512, 0.666 > 1, 0, 1, 512, 1.193 > 1, 1, 1, 512, 0.872 > 2, 0, 0, 512, 0.698 > 2, 2, 0, 512, 0.687 > 2, 0, 2, 512, 1.393 > 2, 2, 2, 512, 0.944 > 3, 0, 0, 512, 0.691 > 3, 3, 0, 512, 0.676 > 3, 0, 3, 512, 1.388 > 3, 3, 3, 512, 0.948 > 4, 0, 0, 512, 0.74 > 4, 4, 0, 512, 0.678 > 4, 0, 4, 512, 1.421 > 4, 4, 4, 512, 0.943 > 5, 0, 0, 512, 0.691 > 5, 5, 0, 512, 0.675 > 5, 0, 5, 512, 1.348 > 5, 5, 5, 512, 0.952 > 6, 0, 0, 512, 0.685 > 6, 6, 0, 512, 0.67 > 6, 0, 6, 512, 1.333 > 6, 6, 6, 512, 0.95 > 7, 0, 0, 512, 0.688 > 7, 7, 0, 512, 0.675 > 7, 0, 7, 512, 1.344 > 7, 7, 7, 512, 0.919 > 8, 0, 0, 512, 0.716 > 8, 0, 8, 512, 0.935 > 9, 0, 0, 512, 0.716 > 9, 1, 0, 512, 0.712 > 9, 0, 9, 512, 0.956 > 9, 1, 9, 512, 0.992 > 10, 0, 0, 512, 0.699 > 10, 2, 0, 512, 0.68 > 10, 0, 10, 512, 0.952 > 10, 2, 10, 512, 0.932 > 11, 0, 0, 512, 0.705 > 11, 3, 0, 512, 0.685 > 11, 0, 11, 512, 0.956 > 11, 3, 11, 512, 0.927 > 12, 0, 0, 512, 0.695 > 12, 4, 0, 512, 0.675 > 12, 0, 12, 512, 0.948 > 12, 4, 12, 512, 0.928 > 13, 0, 0, 512, 0.7 > 13, 5, 0, 512, 0.678 > 13, 0, 13, 512, 0.944 > 13, 5, 13, 512, 0.931 > 14, 0, 0, 512, 0.703 > 14, 6, 0, 512, 0.678 > 14, 0, 14, 512, 0.949 > 14, 6, 14, 512, 0.93 > 15, 0, 0, 512, 0.694 > 15, 7, 0, 512, 0.678 > 15, 0, 15, 512, 0.953 > 15, 7, 15, 512, 0.924 > 16, 0, 0, 512, 1.021 > 16, 0, 16, 512, 1.067 > 17, 0, 0, 512, 0.991 > 17, 1, 0, 512, 0.984 > 17, 0, 17, 512, 0.979 > 17, 1, 17, 512, 0.993 > 18, 0, 0, 512, 0.992 > 18, 2, 0, 512, 1.008 > 18, 0, 18, 512, 1.016 > 18, 2, 18, 512, 0.993 > 19, 0, 0, 512, 0.984 > 19, 3, 0, 512, 0.985 > 19, 0, 19, 512, 1.007 > 19, 3, 19, 512, 1.006 > 20, 0, 0, 512, 0.969 > 20, 4, 0, 512, 0.968 > 20, 0, 20, 512, 0.975 > 20, 4, 20, 512, 0.975 > 21, 0, 0, 512, 0.992 > 21, 5, 0, 512, 0.992 > 21, 0, 21, 512, 0.98 > 21, 5, 21, 512, 0.97 > 22, 0, 0, 512, 0.989 > 22, 6, 0, 512, 0.987 > 22, 0, 22, 512, 0.99 > 22, 6, 22, 512, 0.985 > 23, 0, 0, 512, 0.989 > 23, 7, 0, 512, 0.98 > 23, 0, 23, 512, 1.0 > 23, 7, 23, 512, 0.993 > 24, 0, 0, 512, 0.99 > 24, 0, 24, 512, 0.998 > 25, 0, 0, 512, 1.01 > 25, 1, 0, 512, 1.0 > 25, 0, 25, 512, 0.97 > 25, 1, 25, 512, 0.967 > 26, 0, 0, 512, 1.009 > 26, 2, 0, 512, 0.986 > 26, 0, 26, 512, 0.997 > 26, 2, 26, 512, 0.993 > 27, 0, 0, 512, 0.984 > 27, 3, 0, 512, 0.997 > 27, 0, 27, 512, 0.989 > 27, 3, 27, 512, 0.976 > 28, 0, 0, 512, 0.991 > 28, 4, 0, 512, 1.003 > 28, 0, 28, 512, 0.986 > 28, 4, 28, 512, 0.989 > 29, 0, 0, 512, 0.986 > 29, 5, 0, 512, 0.985 > 29, 0, 29, 512, 0.984 > 29, 5, 29, 512, 0.977 > 30, 0, 0, 512, 0.991 > 30, 6, 0, 512, 0.987 > 30, 0, 30, 512, 0.979 > 30, 6, 30, 512, 0.974 > 31, 0, 0, 512, 0.995 > 31, 7, 0, 512, 0.995 > 31, 0, 31, 512, 0.994 > 31, 7, 31, 512, 0.984 > 4, 0, 0, 32, 0.861 > 4, 1, 0, 32, 0.864 > 4, 0, 1, 32, 0.962 > 4, 1, 1, 32, 0.967 > 4, 0, 0, 64, 0.884 > 4, 2, 0, 64, 0.818 > 4, 0, 2, 64, 0.889 > 4, 2, 2, 64, 0.918 > 4, 0, 0, 128, 0.942 > 4, 3, 0, 128, 0.884 > 4, 0, 3, 128, 0.931 > 4, 3, 3, 128, 0.883 > 4, 0, 0, 256, 0.964 > 4, 4, 0, 256, 0.922 > 4, 0, 4, 256, 0.956 > 4, 4, 4, 256, 0.93 > 4, 5, 0, 512, 0.833 > 4, 0, 5, 512, 1.027 > 4, 5, 5, 512, 0.929 > 4, 0, 0, 1024, 0.998 > 4, 6, 0, 1024, 0.986 > 4, 0, 6, 1024, 0.984 > 4, 6, 6, 1024, 0.977 > 4, 0, 0, 2048, 0.991 > 4, 7, 0, 2048, 0.987 > 4, 0, 7, 2048, 0.996 > 4, 7, 7, 2048, 0.98 > 10, 1, 0, 64, 0.826 > 10, 1, 1, 64, 0.907 > 10, 2, 0, 64, 0.829 > 10, 2, 2, 64, 0.91 > 10, 3, 0, 64, 0.83 > 10, 3, 3, 64, 0.915 > 10, 4, 0, 64, 0.83 > 10, 4, 4, 64, 0.911 > 10, 5, 0, 64, 0.828 > 10, 5, 5, 64, 0.905 > 10, 6, 0, 64, 0.828 > 10, 6, 6, 64, 0.812 > 10, 7, 0, 64, 0.83 > 10, 7, 7, 64, 0.819 > 6, 0, 0, 0, 1.261 > 6, 0, 0, 1, 1.252 > 6, 0, 1, 1, 0.845 > 6, 0, 0, 2, 1.27 > 6, 0, 2, 2, 0.85 > 6, 0, 0, 3, 1.269 > 6, 0, 3, 3, 0.845 > 6, 0, 0, 4, 1.287 > 6, 0, 4, 4, 0.852 > 6, 0, 0, 5, 1.278 > 6, 0, 5, 5, 0.851 > 6, 0, 0, 6, 1.269 > 6, 0, 6, 6, 0.841 > 6, 0, 0, 7, 1.268 > 6, 0, 7, 7, 0.851 > 6, 0, 0, 8, 1.291 > 6, 0, 8, 8, 0.837 > 6, 0, 0, 9, 1.283 > 6, 0, 9, 9, 0.831 > 6, 0, 0, 10, 1.252 > 6, 0, 10, 10, 0.997 > 6, 0, 0, 11, 1.295 > 6, 0, 11, 11, 1.046 > 6, 0, 0, 12, 1.296 > 6, 0, 12, 12, 1.038 > 6, 0, 0, 13, 1.287 > 6, 0, 13, 13, 1.082 > 6, 0, 0, 14, 1.284 > 6, 0, 14, 14, 1.001 > 6, 0, 0, 15, 1.286 > 6, 0, 15, 15, 1.002 > 6, 0, 0, 16, 0.894 > 6, 0, 16, 16, 0.874 > 6, 0, 0, 17, 0.892 > 6, 0, 17, 17, 0.974 > 6, 0, 0, 18, 0.907 > 6, 0, 18, 18, 0.993 > 6, 0, 0, 19, 0.909 > 6, 0, 19, 19, 0.99 > 6, 0, 0, 20, 0.894 > 6, 0, 20, 20, 0.978 > 6, 0, 0, 21, 0.89 > 6, 0, 21, 21, 0.958 > 6, 0, 0, 22, 0.893 > 6, 0, 22, 22, 0.99 > 6, 0, 0, 23, 0.899 > 6, 0, 23, 23, 0.986 > 6, 0, 0, 24, 0.893 > 6, 0, 24, 24, 0.989 > 6, 0, 0, 25, 0.889 > 6, 0, 25, 25, 0.982 > 6, 0, 0, 26, 0.889 > 6, 0, 26, 26, 0.852 > 6, 0, 0, 27, 0.89 > 6, 0, 27, 27, 0.832 > 6, 0, 0, 28, 0.89 > 6, 0, 28, 28, 0.831 > 6, 0, 0, 29, 0.89 > 6, 0, 29, 29, 0.838 > 6, 0, 0, 30, 0.907 > 6, 0, 30, 30, 0.833 > 6, 0, 0, 31, 0.888 > 6, 0, 31, 31, 0.837 > 6, 0, 0, 32, 0.853 > 6, 0, 32, 32, 0.828 > 6, 0, 0, 33, 0.857 > 6, 0, 33, 33, 0.947 > 6, 0, 0, 34, 0.847 > 6, 0, 34, 34, 0.954 > 6, 0, 0, 35, 0.841 > 6, 0, 35, 35, 0.94 > 6, 0, 0, 36, 0.854 > 6, 0, 36, 36, 0.958 > 6, 0, 0, 37, 0.856 > 6, 0, 37, 37, 0.957 > 6, 0, 0, 38, 0.839 > 6, 0, 38, 38, 0.962 > 6, 0, 0, 39, 0.866 > 6, 0, 39, 39, 0.945 > 6, 0, 0, 40, 0.845 > 6, 0, 40, 40, 0.961 > 6, 0, 0, 41, 0.858 > 6, 0, 41, 41, 0.961 > 6, 0, 0, 42, 0.862 > 6, 0, 42, 42, 0.825 > 6, 0, 0, 43, 0.864 > 6, 0, 43, 43, 0.82 > 6, 0, 0, 44, 0.843 > 6, 0, 44, 44, 0.81 > 6, 0, 0, 45, 0.859 > 6, 0, 45, 45, 0.816 > 6, 0, 0, 46, 0.866 > 6, 0, 46, 46, 0.81 > 6, 0, 0, 47, 0.858 > 6, 0, 47, 47, 0.807 > 6, 0, 0, 48, 0.87 > 6, 0, 48, 48, 0.87 > 6, 0, 0, 49, 0.871 > 6, 0, 49, 49, 0.874 > 6, 0, 0, 50, 0.87 > 6, 0, 50, 50, 0.881 > 6, 0, 0, 51, 0.868 > 6, 0, 51, 51, 0.875 > 6, 0, 0, 52, 0.873 > 6, 0, 52, 52, 0.871 > 6, 0, 0, 53, 0.866 > 6, 0, 53, 53, 0.882 > 6, 0, 0, 54, 0.863 > 6, 0, 54, 54, 0.876 > 6, 0, 0, 55, 0.851 > 6, 0, 55, 55, 0.871 > 6, 0, 0, 56, 0.867 > 6, 0, 56, 56, 0.888 > 6, 0, 0, 57, 0.862 > 6, 0, 57, 57, 0.899 > 6, 0, 0, 58, 0.873 > 6, 0, 58, 58, 0.798 > 6, 0, 0, 59, 0.881 > 6, 0, 59, 59, 0.785 > 6, 0, 0, 60, 0.867 > 6, 0, 60, 60, 0.797 > 6, 0, 0, 61, 0.872 > 6, 0, 61, 61, 0.791 > 6, 0, 0, 62, 0.859 > 6, 0, 62, 62, 0.79 > 6, 0, 0, 63, 0.87 > 6, 0, 63, 63, 0.796 > > sysdeps/x86_64/multiarch/strspn-c.c | 86 +++++++++++++---------------- > 1 file changed, 39 insertions(+), 47 deletions(-) > > diff --git a/sysdeps/x86_64/multiarch/strspn-c.c b/sysdeps/x86_64/multiarch/strspn-c.c > index 8fb3aba64d..6124033ceb 100644 > --- a/sysdeps/x86_64/multiarch/strspn-c.c > +++ b/sysdeps/x86_64/multiarch/strspn-c.c > @@ -62,81 +62,73 @@ __strspn_sse42 (const char *s, const char *a) > return 0; > > const char *aligned; > - __m128i mask; > - int offset = (int) ((size_t) a & 15); > + __m128i mask, maskz, zero; > + unsigned int maskz_bits; > + unsigned int offset = (int) ((size_t) a & 15); > + zero = _mm_set1_epi8 (0); > if (offset != 0) > { > /* Load masks. */ > aligned = (const char *) ((size_t) a & -16L); > __m128i mask0 = _mm_load_si128 ((__m128i *) aligned); > - > - mask = __m128i_shift_right (mask0, offset); > + maskz = _mm_cmpeq_epi8 (mask0, zero); > > /* Find where the NULL terminator is. */ > - int length = _mm_cmpistri (mask, mask, 0x3a); > - if (length == 16 - offset) > - { > - /* There is no NULL terminator. */ > - __m128i mask1 = _mm_load_si128 ((__m128i *) (aligned + 16)); > - int index = _mm_cmpistri (mask1, mask1, 0x3a); > - length += index; > - > - /* Don't use SSE4.2 if the length of A > 16. */ > - if (length > 16) > - return __strspn_sse2 (s, a); > - > - if (index != 0) > - { > - /* Combine mask0 and mask1. We could play games with > - palignr, but frankly this data should be in L1 now > - so do the merge via an unaligned load. */ > - mask = _mm_loadu_si128 ((__m128i *) a); > - } > - } > + maskz_bits = _mm_movemask_epi8 (maskz) >> offset; > + if (maskz_bits != 0) > + { > + mask = __m128i_shift_right (mask0, offset); > + offset = (unsigned int) ((size_t) s & 15); > + if (offset) > + goto start_unaligned; > + > + aligned = s; > + goto start_loop; > + } > } > - else > - { > - /* A is aligned. */ > - mask = _mm_load_si128 ((__m128i *) a); > > - /* Find where the NULL terminator is. */ > - int length = _mm_cmpistri (mask, mask, 0x3a); > - if (length == 16) > - { > - /* There is no NULL terminator. Don't use SSE4.2 if the length > - of A > 16. */ > - if (a[16] != 0) > - return __strspn_sse2 (s, a); > - } > + /* A is aligned. */ > + mask = _mm_loadu_si128 ((__m128i *) a); > + > + /* Find where the NULL terminator is. */ > + maskz = _mm_cmpeq_epi8 (mask, zero); > + maskz_bits = _mm_movemask_epi8 (maskz); > + if (maskz_bits == 0) > + { > + /* There is no NULL terminator. Don't use SSE4.2 if the length > + of A > 16. */ > + if (a[16] != 0) > + return __strspn_sse2 (s, a); > } > + aligned = s; > + offset = (unsigned int) ((size_t) s & 15); > > - offset = (int) ((size_t) s & 15); > if (offset != 0) > { > + start_unaligned: > /* Check partial string. */ > aligned = (const char *) ((size_t) s & -16L); > __m128i value = _mm_load_si128 ((__m128i *) aligned); > + __m128i adj_value = __m128i_shift_right (value, offset); > > - value = __m128i_shift_right (value, offset); > - > - int length = _mm_cmpistri (mask, value, 0x12); > + unsigned int length = _mm_cmpistri (mask, adj_value, 0x12); > /* No need to check CFlag since it is always 1. */ > if (length < 16 - offset) > return length; > /* Find where the NULL terminator is. */ > - int index = _mm_cmpistri (value, value, 0x3a); > - if (index < 16 - offset) > + maskz = _mm_cmpeq_epi8 (value, zero); > + maskz_bits = _mm_movemask_epi8 (maskz) >> offset; > + if (maskz_bits != 0) > return length; > aligned += 16; > } > - else > - aligned = s; > > +start_loop: > while (1) > { > __m128i value = _mm_load_si128 ((__m128i *) aligned); > - int index = _mm_cmpistri (mask, value, 0x12); > - int cflag = _mm_cmpistrc (mask, value, 0x12); > + unsigned int index = _mm_cmpistri (mask, value, 0x12); > + unsigned int cflag = _mm_cmpistrc (mask, value, 0x12); > if (cflag) > return (size_t) (aligned + index - s); > aligned += 16; > -- > 2.25.1 > LGTM. Reviewed-by: H.J. Lu Thanks. -- H.J.