From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pj1-x102f.google.com (mail-pj1-x102f.google.com [IPv6:2607:f8b0:4864:20::102f]) by sourceware.org (Postfix) with ESMTPS id E38A93857C7D for ; Tue, 12 Jul 2022 19:07:47 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org E38A93857C7D Received: by mail-pj1-x102f.google.com with SMTP id s21so8766096pjq.4 for ; Tue, 12 Jul 2022 12:07:47 -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=BI38XsGgAW9UNY7WjXaexViAe+LLx8wUuiKwLL7EscI=; b=MLEe1nb8s3fLEd+yM2uv272+/b8O+YOV278Lh88cz++B3KmPQ/XbrqWUDshelvLsxH nxntMH/39dz/BmZpBQX5AkdVU+ssYFg+h7TSyJnDFnKvfbwJlSWVYQ66ff7KsFwnFggv LI9SyI7EV4iBBMKmbie1c7UAR2yEs5SLkLpxQ3cnJDquM16xkaJIPFKHCxOi5+0zqs0m nFcpgX1D6c0MvvjlTF/E6RDqrbXgwOj0wCzsgwvtzM+6lBRxwvvm6k/GT9WCmJKkJSD0 kVXkgPcRrKeRuJ3rAfyakhqX5a4cJq9Ch7mK9de1vK8iavFVcVWQnLsZlxCHKNpXVpL3 libA== X-Gm-Message-State: AJIora/Z3fBxsAk5U5cYqi8szyHG7BcYAoBAEjULgHzCikkS9ezDrWB4 TsaFhIev3zoUpw3OYdWcgVS4VxgLQHY7pSv+PVZlsXxF X-Google-Smtp-Source: AGRyM1uk2iAfJwOm2uFwcJuB2jBUX5oID1sqcPBGgLSFh/Lf4LAx1EiPLD4Jwuve5UvMlS3iqmyBJHw7ZobkRfX/B7Q= X-Received: by 2002:a17:90a:eacd:b0:1ef:84c2:418d with SMTP id ev13-20020a17090aeacd00b001ef84c2418dmr5870885pjb.101.1657652866924; Tue, 12 Jul 2022 12:07:46 -0700 (PDT) MIME-Version: 1.0 References: <20220711220730.1968923-1-goldstein.w.n@gmail.com> <20220712184804.2374822-1-goldstein.w.n@gmail.com> In-Reply-To: <20220712184804.2374822-1-goldstein.w.n@gmail.com> From: "H.J. Lu" Date: Tue, 12 Jul 2022 12:07:11 -0700 Message-ID: Subject: Re: [PATCH v3] x86: Remove __mmask intrinsics in strstr-avx512.c To: Noah Goldstein Cc: GNU C Library , "Carlos O'Donell" Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3024.7 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=3.4.6 X-Spam-Checker-Version: SpamAssassin 3.4.6 (2021-04-09) 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: Tue, 12 Jul 2022 19:07:50 -0000 On Tue, Jul 12, 2022 at 11:48 AM Noah Goldstein wrote: > > The intrinsics are not available before GCC7 and using standard > operators generates code of equivilent or better quality. generate equivalent Otherwise LGTM. Thanks. > > Removed: > _cvtmask64_u64 > _kshiftri_mask64 > _kand_mask64 > > Geometric Mean of 5 Runs of Full Benchmark Suite New / Old: 0.958 > --- > sysdeps/x86_64/multiarch/strstr-avx512.c | 16 ++++++++++------ > 1 file changed, 10 insertions(+), 6 deletions(-) > > diff --git a/sysdeps/x86_64/multiarch/strstr-avx512.c b/sysdeps/x86_64/multiarch/strstr-avx512.c > index 2ab9e96db8..e44c1a05dc 100644 > --- a/sysdeps/x86_64/multiarch/strstr-avx512.c > +++ b/sysdeps/x86_64/multiarch/strstr-avx512.c > @@ -26,6 +26,10 @@ > #define ZMM_SIZE_IN_BYTES 64 > #define PAGESIZE 4096 > > +#define cvtmask64_u64(...) (uint64_t) (__VA_ARGS__) > +#define kshiftri_mask64(x, y) ((x) >> (y)) > +#define kand_mask64(x, y) ((x) & (y)) > + > /* > Returns the index of the first edge within the needle, returns 0 if no edge > is found. Example: 'ab' is the first edge in 'aaaaaaaaaabaarddg' > @@ -133,15 +137,15 @@ __strstr_avx512 (const char *haystack, const char *ned) > __m512i hay0 = _mm512_maskz_loadu_epi8 (loadmask, haystack + hay_index); > /* Search for NULL and compare only till null char */ > uint64_t nullmask > - = _cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0)); > + = cvtmask64_u64 (_mm512_mask_testn_epi8_mask (loadmask, hay0, hay0)); > uint64_t cmpmask = nullmask ^ (nullmask - ONE_64BIT); > - cmpmask = cmpmask & _cvtmask64_u64 (loadmask); > + cmpmask = cmpmask & cvtmask64_u64 (loadmask); > /* Search for the 2 charaters of needle */ > __mmask64 k0 = _mm512_cmpeq_epi8_mask (hay0, ned0); > __mmask64 k1 = _mm512_cmpeq_epi8_mask (hay0, ned1); > - k1 = _kshiftri_mask64 (k1, 1); > + k1 = kshiftri_mask64 (k1, 1); > /* k2 masks tell us if both chars from needle match */ > - uint64_t k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask; > + uint64_t k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask; > /* For every match, search for the entire needle for a full match */ > while (k2) > { > @@ -178,13 +182,13 @@ __strstr_avx512 (const char *haystack, const char *ned) > hay0 = _mm512_loadu_si512 (haystack + hay_index); > hay1 = _mm512_load_si512 (haystack + hay_index > + 1); // Always 64 byte aligned > - nullmask = _cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1)); > + nullmask = cvtmask64_u64 (_mm512_testn_epi8_mask (hay1, hay1)); > /* Compare only till null char */ > cmpmask = nullmask ^ (nullmask - ONE_64BIT); > k0 = _mm512_cmpeq_epi8_mask (hay0, ned0); > k1 = _mm512_cmpeq_epi8_mask (hay1, ned1); > /* k2 masks tell us if both chars from needle match */ > - k2 = _cvtmask64_u64 (_kand_mask64 (k0, k1)) & cmpmask; > + k2 = cvtmask64_u64 (kand_mask64 (k0, k1)) & cmpmask; > /* For every match, compare full strings for potential match */ > while (k2) > { > -- > 2.34.1 > -- H.J.