From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf1-x430.google.com (mail-pf1-x430.google.com [IPv6:2607:f8b0:4864:20::430]) by sourceware.org (Postfix) with ESMTPS id BEC42385741A for ; Thu, 24 Mar 2022 18:56:14 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org BEC42385741A Received: by mail-pf1-x430.google.com with SMTP id t5so4669330pfg.4 for ; Thu, 24 Mar 2022 11:56:14 -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=BrlvjTh0nNU1zRCzoyTkvfKolIF4I1ZfMiYJSshx/b0=; b=lJBJTV9dFOrpNW7ZrNB0j5vs6eTFUMwfq56mm28xg8aI7SSTT886KwRME78sVTXrN9 jEWIc3Kzyei2rqKDet8ioSb7ihK7Ckv8pM1lhUotoXFqYGrjHRmOku3+aqSCIdNHt+T/ NqsJmEhRWwa2PsNEutv3u75FxB8wnzHEONbXIrdJ+NRynNezjSa8EeK1h/twPgPNERSI c/euTVQY1oh5ffVS9QtQUctYV6wUel83VXz6mK0ydE3S7a1WvygWiPN4B6RdonKb60pd Z89dVgW28D2lowniy47+yd3Vko3WEkut/gfUNUJuIdA5Oce6RhQ2C7fO2eK2POFaJDHd qASg== X-Gm-Message-State: AOAM530d3CEo4WekmiaBFnWenQMJDhqa5lF9m//m375irl2YGTUbHr0C 9qMzgSY3BOuZ8CpzFOKvOFzzAdjbU46R23MGWqcAi+UblKs= X-Google-Smtp-Source: ABdhPJxOVhvjQ02PeXN0FkttwHuRKstiXmpHDjp+ilsV0VRmWaLJrBwpXwEUxhxPPaaRfIAOBIkLq96CfetY7soXvnc= X-Received: by 2002:a63:af06:0:b0:378:3582:a49f with SMTP id w6-20020a63af06000000b003783582a49fmr5051381pge.125.1648148173605; Thu, 24 Mar 2022 11:56:13 -0700 (PDT) MIME-Version: 1.0 References: <20220323215734.3927131-1-goldstein.w.n@gmail.com> <20220323215734.3927131-7-goldstein.w.n@gmail.com> In-Reply-To: <20220323215734.3927131-7-goldstein.w.n@gmail.com> From: "H.J. Lu" Date: Thu, 24 Mar 2022 11:55:37 -0700 Message-ID: Subject: Re: [PATCH v1 07/23] x86: Optimize strcspn and strpbrk in strcspn-c.c To: Noah Goldstein Cc: GNU C Library , "Carlos O'Donell" Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3026.1 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 18:56:17 -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/strlen; New / Original: .928 > > All string/memory tests pass. > --- > Geomtric Mean N=20 runs; All functions page aligned > len, align1, align2, pos, New Time / Old Time > 0, 0, 0, 512, 1.207 > 1, 0, 0, 512, 1.039 > 1, 1, 0, 512, 0.997 > 1, 0, 1, 512, 0.981 > 1, 1, 1, 512, 0.977 > 2, 0, 0, 512, 1.02 > 2, 2, 0, 512, 0.979 > 2, 0, 2, 512, 0.902 > 2, 2, 2, 512, 0.958 > 3, 0, 0, 512, 0.978 > 3, 3, 0, 512, 0.988 > 3, 0, 3, 512, 0.979 > 3, 3, 3, 512, 0.955 > 4, 0, 0, 512, 0.969 > 4, 4, 0, 512, 0.991 > 4, 0, 4, 512, 0.94 > 4, 4, 4, 512, 0.958 > 5, 0, 0, 512, 0.963 > 5, 5, 0, 512, 1.004 > 5, 0, 5, 512, 0.948 > 5, 5, 5, 512, 0.971 > 6, 0, 0, 512, 0.933 > 6, 6, 0, 512, 1.007 > 6, 0, 6, 512, 0.921 > 6, 6, 6, 512, 0.969 > 7, 0, 0, 512, 0.928 > 7, 7, 0, 512, 0.976 > 7, 0, 7, 512, 0.932 > 7, 7, 7, 512, 0.995 > 8, 0, 0, 512, 0.931 > 8, 0, 8, 512, 0.766 > 9, 0, 0, 512, 0.965 > 9, 1, 0, 512, 0.999 > 9, 0, 9, 512, 0.765 > 9, 1, 9, 512, 0.97 > 10, 0, 0, 512, 0.976 > 10, 2, 0, 512, 0.991 > 10, 0, 10, 512, 0.768 > 10, 2, 10, 512, 0.926 > 11, 0, 0, 512, 0.958 > 11, 3, 0, 512, 1.006 > 11, 0, 11, 512, 0.768 > 11, 3, 11, 512, 0.908 > 12, 0, 0, 512, 0.945 > 12, 4, 0, 512, 0.896 > 12, 0, 12, 512, 0.764 > 12, 4, 12, 512, 0.785 > 13, 0, 0, 512, 0.957 > 13, 5, 0, 512, 1.019 > 13, 0, 13, 512, 0.76 > 13, 5, 13, 512, 0.785 > 14, 0, 0, 512, 0.918 > 14, 6, 0, 512, 1.004 > 14, 0, 14, 512, 0.78 > 14, 6, 14, 512, 0.711 > 15, 0, 0, 512, 0.855 > 15, 7, 0, 512, 0.985 > 15, 0, 15, 512, 0.779 > 15, 7, 15, 512, 0.772 > 16, 0, 0, 512, 0.987 > 16, 0, 16, 512, 0.99 > 17, 0, 0, 512, 0.996 > 17, 1, 0, 512, 0.979 > 17, 0, 17, 512, 1.001 > 17, 1, 17, 512, 1.03 > 18, 0, 0, 512, 0.976 > 18, 2, 0, 512, 0.989 > 18, 0, 18, 512, 0.976 > 18, 2, 18, 512, 0.992 > 19, 0, 0, 512, 0.991 > 19, 3, 0, 512, 0.988 > 19, 0, 19, 512, 1.009 > 19, 3, 19, 512, 1.018 > 20, 0, 0, 512, 0.999 > 20, 4, 0, 512, 1.005 > 20, 0, 20, 512, 0.993 > 20, 4, 20, 512, 0.983 > 21, 0, 0, 512, 0.982 > 21, 5, 0, 512, 0.988 > 21, 0, 21, 512, 0.978 > 21, 5, 21, 512, 0.984 > 22, 0, 0, 512, 0.988 > 22, 6, 0, 512, 0.979 > 22, 0, 22, 512, 0.984 > 22, 6, 22, 512, 0.983 > 23, 0, 0, 512, 0.996 > 23, 7, 0, 512, 0.998 > 23, 0, 23, 512, 0.979 > 23, 7, 23, 512, 0.987 > 24, 0, 0, 512, 0.99 > 24, 0, 24, 512, 0.979 > 25, 0, 0, 512, 0.985 > 25, 1, 0, 512, 0.988 > 25, 0, 25, 512, 0.99 > 25, 1, 25, 512, 0.986 > 26, 0, 0, 512, 1.005 > 26, 2, 0, 512, 0.995 > 26, 0, 26, 512, 0.992 > 26, 2, 26, 512, 0.983 > 27, 0, 0, 512, 0.986 > 27, 3, 0, 512, 0.978 > 27, 0, 27, 512, 0.986 > 27, 3, 27, 512, 0.973 > 28, 0, 0, 512, 0.995 > 28, 4, 0, 512, 0.993 > 28, 0, 28, 512, 0.983 > 28, 4, 28, 512, 1.005 > 29, 0, 0, 512, 0.983 > 29, 5, 0, 512, 0.982 > 29, 0, 29, 512, 0.984 > 29, 5, 29, 512, 1.005 > 30, 0, 0, 512, 0.978 > 30, 6, 0, 512, 0.985 > 30, 0, 30, 512, 0.994 > 30, 6, 30, 512, 0.993 > 31, 0, 0, 512, 0.984 > 31, 7, 0, 512, 0.983 > 31, 0, 31, 512, 1.0 > 31, 7, 31, 512, 1.031 > 4, 0, 0, 32, 0.916 > 4, 1, 0, 32, 0.952 > 4, 0, 1, 32, 0.927 > 4, 1, 1, 32, 0.969 > 4, 0, 0, 64, 0.961 > 4, 2, 0, 64, 0.955 > 4, 0, 2, 64, 0.975 > 4, 2, 2, 64, 0.972 > 4, 0, 0, 128, 0.971 > 4, 3, 0, 128, 0.982 > 4, 0, 3, 128, 0.945 > 4, 3, 3, 128, 0.971 > 4, 0, 0, 256, 1.004 > 4, 4, 0, 256, 0.966 > 4, 0, 4, 256, 0.961 > 4, 4, 4, 256, 0.971 > 4, 5, 0, 512, 0.929 > 4, 0, 5, 512, 0.969 > 4, 5, 5, 512, 0.985 > 4, 0, 0, 1024, 1.003 > 4, 6, 0, 1024, 1.009 > 4, 0, 6, 1024, 1.005 > 4, 6, 6, 1024, 0.999 > 4, 0, 0, 2048, 0.917 > 4, 7, 0, 2048, 1.015 > 4, 0, 7, 2048, 1.011 > 4, 7, 7, 2048, 0.907 > 10, 1, 0, 64, 0.964 > 10, 1, 1, 64, 0.966 > 10, 2, 0, 64, 0.953 > 10, 2, 2, 64, 0.972 > 10, 3, 0, 64, 0.962 > 10, 3, 3, 64, 0.969 > 10, 4, 0, 64, 0.957 > 10, 4, 4, 64, 0.969 > 10, 5, 0, 64, 0.961 > 10, 5, 5, 64, 0.965 > 10, 6, 0, 64, 0.949 > 10, 6, 6, 64, 0.9 > 10, 7, 0, 64, 0.957 > 10, 7, 7, 64, 0.897 > 6, 0, 0, 0, 0.991 > 6, 0, 0, 1, 1.011 > 6, 0, 1, 1, 0.939 > 6, 0, 0, 2, 1.016 > 6, 0, 2, 2, 0.94 > 6, 0, 0, 3, 1.019 > 6, 0, 3, 3, 0.941 > 6, 0, 0, 4, 1.056 > 6, 0, 4, 4, 0.884 > 6, 0, 0, 5, 0.977 > 6, 0, 5, 5, 0.934 > 6, 0, 0, 6, 0.954 > 6, 0, 6, 6, 0.93 > 6, 0, 0, 7, 0.963 > 6, 0, 7, 7, 0.916 > 6, 0, 0, 8, 0.963 > 6, 0, 8, 8, 0.945 > 6, 0, 0, 9, 1.028 > 6, 0, 9, 9, 0.942 > 6, 0, 0, 10, 0.955 > 6, 0, 10, 10, 0.831 > 6, 0, 0, 11, 0.948 > 6, 0, 11, 11, 0.82 > 6, 0, 0, 12, 1.033 > 6, 0, 12, 12, 0.873 > 6, 0, 0, 13, 0.983 > 6, 0, 13, 13, 0.852 > 6, 0, 0, 14, 0.984 > 6, 0, 14, 14, 0.853 > 6, 0, 0, 15, 0.984 > 6, 0, 15, 15, 0.882 > 6, 0, 0, 16, 0.971 > 6, 0, 16, 16, 0.958 > 6, 0, 0, 17, 0.938 > 6, 0, 17, 17, 0.947 > 6, 0, 0, 18, 0.96 > 6, 0, 18, 18, 0.938 > 6, 0, 0, 19, 0.903 > 6, 0, 19, 19, 0.943 > 6, 0, 0, 20, 0.947 > 6, 0, 20, 20, 0.951 > 6, 0, 0, 21, 0.948 > 6, 0, 21, 21, 0.96 > 6, 0, 0, 22, 0.926 > 6, 0, 22, 22, 0.951 > 6, 0, 0, 23, 0.923 > 6, 0, 23, 23, 0.959 > 6, 0, 0, 24, 0.918 > 6, 0, 24, 24, 0.952 > 6, 0, 0, 25, 0.97 > 6, 0, 25, 25, 0.952 > 6, 0, 0, 26, 0.871 > 6, 0, 26, 26, 0.869 > 6, 0, 0, 27, 0.935 > 6, 0, 27, 27, 0.836 > 6, 0, 0, 28, 0.936 > 6, 0, 28, 28, 0.857 > 6, 0, 0, 29, 0.876 > 6, 0, 29, 29, 0.859 > 6, 0, 0, 30, 0.934 > 6, 0, 30, 30, 0.857 > 6, 0, 0, 31, 0.962 > 6, 0, 31, 31, 0.86 > 6, 0, 0, 32, 0.912 > 6, 0, 32, 32, 0.94 > 6, 0, 0, 33, 0.903 > 6, 0, 33, 33, 0.968 > 6, 0, 0, 34, 0.913 > 6, 0, 34, 34, 0.896 > 6, 0, 0, 35, 0.904 > 6, 0, 35, 35, 0.913 > 6, 0, 0, 36, 0.905 > 6, 0, 36, 36, 0.907 > 6, 0, 0, 37, 0.899 > 6, 0, 37, 37, 0.9 > 6, 0, 0, 38, 0.912 > 6, 0, 38, 38, 0.919 > 6, 0, 0, 39, 0.925 > 6, 0, 39, 39, 0.927 > 6, 0, 0, 40, 0.923 > 6, 0, 40, 40, 0.972 > 6, 0, 0, 41, 0.92 > 6, 0, 41, 41, 0.966 > 6, 0, 0, 42, 0.915 > 6, 0, 42, 42, 0.834 > 6, 0, 0, 43, 0.92 > 6, 0, 43, 43, 0.856 > 6, 0, 0, 44, 0.908 > 6, 0, 44, 44, 0.858 > 6, 0, 0, 45, 0.932 > 6, 0, 45, 45, 0.847 > 6, 0, 0, 46, 0.927 > 6, 0, 46, 46, 0.859 > 6, 0, 0, 47, 0.902 > 6, 0, 47, 47, 0.855 > 6, 0, 0, 48, 0.949 > 6, 0, 48, 48, 0.934 > 6, 0, 0, 49, 0.907 > 6, 0, 49, 49, 0.943 > 6, 0, 0, 50, 0.934 > 6, 0, 50, 50, 0.943 > 6, 0, 0, 51, 0.933 > 6, 0, 51, 51, 0.939 > 6, 0, 0, 52, 0.944 > 6, 0, 52, 52, 0.944 > 6, 0, 0, 53, 0.939 > 6, 0, 53, 53, 0.938 > 6, 0, 0, 54, 0.9 > 6, 0, 54, 54, 0.923 > 6, 0, 0, 55, 0.9 > 6, 0, 55, 55, 0.927 > 6, 0, 0, 56, 0.9 > 6, 0, 56, 56, 0.917 > 6, 0, 0, 57, 0.9 > 6, 0, 57, 57, 0.916 > 6, 0, 0, 58, 0.914 > 6, 0, 58, 58, 0.784 > 6, 0, 0, 59, 0.863 > 6, 0, 59, 59, 0.846 > 6, 0, 0, 60, 0.88 > 6, 0, 60, 60, 0.827 > 6, 0, 0, 61, 0.896 > 6, 0, 61, 61, 0.847 > 6, 0, 0, 62, 0.894 > 6, 0, 62, 62, 0.865 > 6, 0, 0, 63, 0.934 > 6, 0, 63, 63, 0.866 > > sysdeps/x86_64/multiarch/strcspn-c.c | 83 +++++++++++++--------------- > 1 file changed, 37 insertions(+), 46 deletions(-) > > diff --git a/sysdeps/x86_64/multiarch/strcspn-c.c b/sysdeps/x86_64/multiarch/strcspn-c.c > index 013aebf797..c312fab8b1 100644 > --- a/sysdeps/x86_64/multiarch/strcspn-c.c > +++ b/sysdeps/x86_64/multiarch/strcspn-c.c > @@ -84,83 +84,74 @@ STRCSPN_SSE42 (const char *s, const char *a) > RETURN (NULL, strlen (s)); > > const char *aligned; > - __m128i mask; > - int offset = (int) ((size_t) a & 15); > + __m128i mask, maskz, zero; > + unsigned int maskz_bits; > + unsigned int offset = (unsigned 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 STRCSPN_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 STRCSPN_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 STRCSPN_SSE2 (s, a); > } > > - offset = (int) ((size_t) s & 15); > + aligned = s; > + offset = (unsigned 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); > > value = __m128i_shift_right (value, offset); > > - int length = _mm_cmpistri (mask, value, 0x2); > + unsigned int length = _mm_cmpistri (mask, value, 0x2); > /* No need to check ZFlag since ZFlag is always 1. */ > - int cflag = _mm_cmpistrc (mask, value, 0x2); > + unsigned int cflag = _mm_cmpistrc (mask, value, 0x2); > if (cflag) > RETURN ((char *) (s + length), length); > /* Find where the NULL terminator is. */ > - int index = _mm_cmpistri (value, value, 0x3a); > + unsigned int index = _mm_cmpistri (value, value, 0x3a); > if (index < 16 - offset) > RETURN (NULL, index); > aligned += 16; > } > - else > - aligned = s; > > +start_loop: > while (1) > { > __m128i value = _mm_load_si128 ((__m128i *) aligned); > - int index = _mm_cmpistri (mask, value, 0x2); > - int cflag = _mm_cmpistrc (mask, value, 0x2); > - int zflag = _mm_cmpistrz (mask, value, 0x2); > + unsigned int index = _mm_cmpistri (mask, value, 0x2); > + unsigned int cflag = _mm_cmpistrc (mask, value, 0x2); > + unsigned int zflag = _mm_cmpistrz (mask, value, 0x2); > if (cflag) > RETURN ((char *) (aligned + index), (size_t) (aligned + index - s)); > if (zflag) > -- > 2.25.1 > LGTM. Reviewed-by: H.J. Lu Thanks. -- H.J.