From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pj1-x102d.google.com (mail-pj1-x102d.google.com [IPv6:2607:f8b0:4864:20::102d]) by sourceware.org (Postfix) with ESMTPS id E44EA385803E for ; Thu, 24 Mar 2022 18:53:44 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.4.1 sourceware.org E44EA385803E Received: by mail-pj1-x102d.google.com with SMTP id o6-20020a17090a9f8600b001c6562049d9so5982002pjp.3 for ; Thu, 24 Mar 2022 11:53:44 -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=Alz4GHMfrKVFoK3VdTffPRB7Uc8aT9IfrIbW71EOOrM=; b=JZDBnYAx9UFzgDo21/64bePXzD4O0DBvYZCUh0YL6FzidFoQfNfPu/DuLonyluXrzG 7T9v3E+Wm/4OP41nGBbzT5SaYiwpsco5W7YgUbHXsfp8vXK1tsmRII/cuGM1AxcpWP/w knQvWZ+wryciY6rK5SYLzJhkn9CiZIb4emxQtemUm4a94Tc5Nw4Fljby0G7Vkikbp49O K68uaCXEP4ck8nHq/m9z95axPBg3BulYA97HhuN3OhtFIkv48hT/6JXXzXNG2/uDs8Bb q63Ghe+297JkAVGbyyEql7LlscMKMwg0yOgC41lQa2W4SEQPweTyMydl3S/R5Wcp/N6O 1NiA== X-Gm-Message-State: AOAM533PBgHEemyHIRFGyV3ptWe3zZ/npa/JvVMxJhOo6oc3/Uxh72AL rsnYYJkHFE6snZki0NmlyAw9Qm8zfJnTNnn9Pu/6PESXFd4= X-Google-Smtp-Source: ABdhPJzc5VyZg4Rt64y6BCqpp2f9OfRiebILKKqY6YXk1/cVC0O50WNeUQyRtDxdyjmTt2U9a8GXHCeNEel8N4IdY0k= X-Received: by 2002:a17:902:d508:b0:154:2f99:a52e with SMTP id b8-20020a170902d50800b001542f99a52emr7243594plg.102.1648148023779; Thu, 24 Mar 2022 11:53:43 -0700 (PDT) MIME-Version: 1.0 References: <20220323215734.3927131-1-goldstein.w.n@gmail.com> <20220323215734.3927131-3-goldstein.w.n@gmail.com> In-Reply-To: <20220323215734.3927131-3-goldstein.w.n@gmail.com> From: "H.J. Lu" Date: Thu, 24 Mar 2022 11:53:07 -0700 Message-ID: Subject: Re: [PATCH v1 03/23] x86: Code cleanup in strchr-avx2 and comment justifying branch To: Noah Goldstein Cc: GNU C Library , "Carlos O'Donell" Content-Type: text/plain; charset="UTF-8" X-Spam-Status: No, score=-3026.3 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 18:53:48 -0000 On Wed, Mar 23, 2022 at 2:58 PM Noah Goldstein wrote: > > Small code cleanup for size: -53 bytes. > > Add comment justifying using a branch to do NULL/non-null return. Do you have followup patches to improve its performance? We are backporting all x86-64 improvements to Intel release branches: https://gitlab.com/x86-glibc/glibc/-/wikis/home Patches without performance improvements are undesirable. > All string/memory tests pass and no regressions in benchtests. > > geometric_mean(N=20) of all benchmarks Original / New: 1.00 > --- > Geomtric Mean N=20 runs; All functions page aligned > length, alignment, pos, rand, seek_char/branch, max_char/perc-zero, New Time / Old Time > 2048, 0, 32, 0, 23, 127, 1.033 > 2048, 1, 32, 0, 23, 127, 1.006 > 2048, 0, 64, 0, 23, 127, 1.02 > 2048, 2, 64, 0, 23, 127, 0.992 > 2048, 0, 128, 0, 23, 127, 0.996 > 2048, 3, 128, 0, 23, 127, 0.966 > 2048, 0, 256, 0, 23, 127, 0.996 > 2048, 4, 256, 0, 23, 127, 0.998 > 2048, 0, 512, 0, 23, 127, 0.991 > 2048, 5, 512, 0, 23, 127, 0.991 > 2048, 0, 1024, 0, 23, 127, 0.993 > 2048, 6, 1024, 0, 23, 127, 0.992 > 2048, 0, 2048, 0, 23, 127, 0.992 > 2048, 7, 2048, 0, 23, 127, 0.976 > 4096, 0, 32, 0, 23, 127, 0.983 > 4096, 1, 32, 0, 23, 127, 0.994 > 4096, 0, 64, 0, 23, 127, 0.968 > 4096, 2, 64, 0, 23, 127, 1.018 > 4096, 0, 128, 0, 23, 127, 0.99 > 4096, 3, 128, 0, 23, 127, 1.001 > 4096, 0, 256, 0, 23, 127, 1.0 > 4096, 4, 256, 0, 23, 127, 1.001 > 4096, 0, 512, 0, 23, 127, 0.989 > 4096, 5, 512, 0, 23, 127, 0.988 > 4096, 0, 1024, 0, 23, 127, 0.994 > 4096, 6, 1024, 0, 23, 127, 0.993 > 4096, 0, 2048, 0, 23, 127, 0.987 > 4096, 7, 2048, 0, 23, 127, 0.996 > 256, 1, 64, 0, 23, 127, 1.004 > 256, 2, 64, 0, 23, 127, 1.004 > 256, 3, 64, 0, 23, 127, 0.992 > 256, 4, 64, 0, 23, 127, 1.001 > 256, 5, 64, 0, 23, 127, 1.001 > 256, 6, 64, 0, 23, 127, 0.998 > 256, 7, 64, 0, 23, 127, 0.994 > 512, 0, 256, 0, 23, 127, 0.999 > 512, 16, 256, 0, 23, 127, 1.002 > 512, 32, 256, 0, 23, 127, 0.994 > 512, 48, 256, 0, 23, 127, 0.991 > 512, 64, 256, 0, 23, 127, 0.994 > 512, 80, 256, 0, 23, 127, 0.994 > 512, 96, 256, 0, 23, 127, 0.996 > 512, 112, 256, 0, 23, 127, 0.999 > 1, 0, 0, 0, 23, 127, 0.978 > 2, 0, 1, 0, 23, 127, 0.981 > 3, 0, 2, 0, 23, 127, 0.993 > 4, 0, 3, 0, 23, 127, 1.004 > 5, 0, 4, 0, 23, 127, 1.002 > 6, 0, 5, 0, 23, 127, 0.991 > 7, 0, 6, 0, 23, 127, 0.99 > 8, 0, 7, 0, 23, 127, 1.012 > 9, 0, 8, 0, 23, 127, 0.994 > 10, 0, 9, 0, 23, 127, 1.003 > 11, 0, 10, 0, 23, 127, 0.999 > 12, 0, 11, 0, 23, 127, 1.007 > 13, 0, 12, 0, 23, 127, 1.0 > 14, 0, 13, 0, 23, 127, 0.997 > 15, 0, 14, 0, 23, 127, 0.996 > 16, 0, 15, 0, 23, 127, 0.993 > 17, 0, 16, 0, 23, 127, 1.002 > 18, 0, 17, 0, 23, 127, 0.997 > 19, 0, 18, 0, 23, 127, 0.998 > 20, 0, 19, 0, 23, 127, 0.994 > 21, 0, 20, 0, 23, 127, 0.99 > 22, 0, 21, 0, 23, 127, 0.992 > 23, 0, 22, 0, 23, 127, 0.996 > 24, 0, 23, 0, 23, 127, 0.991 > 25, 0, 24, 0, 23, 127, 0.997 > 26, 0, 25, 0, 23, 127, 1.011 > 27, 0, 26, 0, 23, 127, 1.013 > 28, 0, 27, 0, 23, 127, 0.996 > 29, 0, 28, 0, 23, 127, 0.993 > 30, 0, 29, 0, 23, 127, 1.009 > 31, 0, 30, 0, 23, 127, 1.009 > 32, 0, 31, 0, 23, 127, 1.008 > 2048, 0, 32, 0, 0, 127, 1.0 > 2048, 1, 32, 0, 0, 127, 1.01 > 2048, 0, 64, 0, 0, 127, 0.997 > 2048, 2, 64, 0, 0, 127, 1.002 > 2048, 0, 128, 0, 0, 127, 0.986 > 2048, 3, 128, 0, 0, 127, 0.997 > 2048, 0, 256, 0, 0, 127, 1.002 > 2048, 4, 256, 0, 0, 127, 0.999 > 2048, 0, 512, 0, 0, 127, 0.991 > 2048, 5, 512, 0, 0, 127, 0.984 > 2048, 0, 1024, 0, 0, 127, 0.994 > 2048, 6, 1024, 0, 0, 127, 0.993 > 2048, 0, 2048, 0, 0, 127, 0.951 > 2048, 7, 2048, 0, 0, 127, 0.989 > 4096, 0, 32, 0, 0, 127, 0.993 > 4096, 1, 32, 0, 0, 127, 0.997 > 4096, 0, 64, 0, 0, 127, 1.004 > 4096, 2, 64, 0, 0, 127, 1.016 > 4096, 0, 128, 0, 0, 127, 0.973 > 4096, 3, 128, 0, 0, 127, 1.001 > 4096, 0, 256, 0, 0, 127, 0.999 > 4096, 4, 256, 0, 0, 127, 0.998 > 4096, 0, 512, 0, 0, 127, 0.99 > 4096, 5, 512, 0, 0, 127, 0.985 > 4096, 0, 1024, 0, 0, 127, 0.993 > 4096, 6, 1024, 0, 0, 127, 0.997 > 4096, 0, 2048, 0, 0, 127, 0.995 > 4096, 7, 2048, 0, 0, 127, 0.996 > 256, 1, 64, 0, 0, 127, 1.01 > 256, 2, 64, 0, 0, 127, 1.024 > 256, 3, 64, 0, 0, 127, 1.03 > 256, 4, 64, 0, 0, 127, 1.004 > 256, 5, 64, 0, 0, 127, 0.998 > 256, 6, 64, 0, 0, 127, 0.998 > 256, 7, 64, 0, 0, 127, 0.997 > 512, 0, 256, 0, 0, 127, 0.996 > 512, 16, 256, 0, 0, 127, 0.995 > 512, 32, 256, 0, 0, 127, 0.996 > 512, 48, 256, 0, 0, 127, 0.992 > 512, 64, 256, 0, 0, 127, 0.999 > 512, 80, 256, 0, 0, 127, 1.002 > 512, 96, 256, 0, 0, 127, 0.999 > 512, 112, 256, 0, 0, 127, 0.998 > 1, 0, 0, 0, 0, 127, 1.016 > 2, 0, 1, 0, 0, 127, 0.998 > 3, 0, 2, 0, 0, 127, 1.02 > 4, 0, 3, 0, 0, 127, 1.004 > 5, 0, 4, 0, 0, 127, 1.021 > 6, 0, 5, 0, 0, 127, 1.014 > 7, 0, 6, 0, 0, 127, 1.007 > 8, 0, 7, 0, 0, 127, 1.016 > 9, 0, 8, 0, 0, 127, 1.003 > 10, 0, 9, 0, 0, 127, 1.004 > 11, 0, 10, 0, 0, 127, 0.995 > 12, 0, 11, 0, 0, 127, 1.009 > 13, 0, 12, 0, 0, 127, 1.005 > 14, 0, 13, 0, 0, 127, 0.987 > 15, 0, 14, 0, 0, 127, 0.998 > 16, 0, 15, 0, 0, 127, 1.004 > 17, 0, 16, 0, 0, 127, 1.01 > 18, 0, 17, 0, 0, 127, 1.01 > 19, 0, 18, 0, 0, 127, 1.006 > 20, 0, 19, 0, 0, 127, 1.012 > 21, 0, 20, 0, 0, 127, 0.999 > 22, 0, 21, 0, 0, 127, 1.004 > 23, 0, 22, 0, 0, 127, 0.988 > 24, 0, 23, 0, 0, 127, 0.993 > 25, 0, 24, 0, 0, 127, 1.004 > 26, 0, 25, 0, 0, 127, 0.99 > 27, 0, 26, 0, 0, 127, 1.016 > 28, 0, 27, 0, 0, 127, 0.987 > 29, 0, 28, 0, 0, 127, 0.989 > 30, 0, 29, 0, 0, 127, 0.998 > 31, 0, 30, 0, 0, 127, 1.005 > 32, 0, 31, 0, 0, 127, 0.993 > > 16, 0, 15, 1, 1, 0, 1.002 > 16, 0, 15, 1, 0, 0, 1.0 > 16, 0, 15, 1, 1, 0.1, 1.034 > 16, 0, 15, 1, 0, 0.1, 1.03 > 16, 0, 15, 1, 1, 0.25, 0.993 > 16, 0, 15, 1, 0, 0.25, 1.081 > 16, 0, 15, 1, 1, 0.33, 0.959 > 16, 0, 15, 1, 0, 0.33, 1.142 > 16, 0, 15, 1, 1, 0.5, 0.929 > 16, 0, 15, 1, 0, 0.5, 1.072 > 16, 0, 15, 1, 1, 0.66, 0.984 > 16, 0, 15, 1, 0, 0.66, 1.069 > 16, 0, 15, 1, 1, 0.75, 0.969 > 16, 0, 15, 1, 0, 0.75, 1.059 > 16, 0, 15, 1, 1, 0.9, 0.98 > 16, 0, 15, 1, 0, 0.9, 0.994 > 16, 0, 15, 1, 1, 1, 0.993 > 16, 0, 15, 1, 0, 1, 0.996 > > sysdeps/x86_64/multiarch/strchr-avx2.S | 204 +++++++++++++------------ > 1 file changed, 107 insertions(+), 97 deletions(-) > > diff --git a/sysdeps/x86_64/multiarch/strchr-avx2.S b/sysdeps/x86_64/multiarch/strchr-avx2.S > index 086cabf76a..1a916cc951 100644 > --- a/sysdeps/x86_64/multiarch/strchr-avx2.S > +++ b/sysdeps/x86_64/multiarch/strchr-avx2.S > @@ -48,13 +48,13 @@ > # define PAGE_SIZE 4096 > > .section SECTION(.text),"ax",@progbits > -ENTRY (STRCHR) > +ENTRY_P2ALIGN (STRCHR, 5) > /* Broadcast CHAR to YMM0. */ > vmovd %esi, %xmm0 > movl %edi, %eax > andl $(PAGE_SIZE - 1), %eax > VPBROADCAST %xmm0, %ymm0 > - vpxor %xmm9, %xmm9, %xmm9 > + vpxor %xmm1, %xmm1, %xmm1 > > /* Check if we cross page boundary with one vector load. */ > cmpl $(PAGE_SIZE - VEC_SIZE), %eax > @@ -62,37 +62,29 @@ ENTRY (STRCHR) > > /* Check the first VEC_SIZE bytes. Search for both CHAR and the > null byte. */ > - vmovdqu (%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqu (%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jz L(aligned_more) > tzcntl %eax, %eax > # ifndef USE_AS_STRCHRNUL > - /* Found CHAR or the null byte. */ > - cmp (%rdi, %rax), %CHAR_REG > - jne L(zero) > -# endif > - addq %rdi, %rax > - VZEROUPPER_RETURN > - > - /* .p2align 5 helps keep performance more consistent if ENTRY() > - alignment % 32 was either 16 or 0. As well this makes the > - alignment % 32 of the loop_4x_vec fixed which makes tuning it > - easier. */ > - .p2align 5 > -L(first_vec_x4): > - tzcntl %eax, %eax > - addq $(VEC_SIZE * 3 + 1), %rdi > -# ifndef USE_AS_STRCHRNUL > - /* Found CHAR or the null byte. */ > + /* Found CHAR or the null byte. */ > cmp (%rdi, %rax), %CHAR_REG > + /* NB: Use a branch instead of cmovcc here. The expectation is > + that with strchr the user will branch based on input being > + null. Since this branch will be 100% predictive of the user > + branch a branch miss here should save what otherwise would > + be branch miss in the user code. Otherwise using a branch 1) > + saves code size and 2) is faster in highly predictable > + environments. */ > jne L(zero) > # endif > addq %rdi, %rax > - VZEROUPPER_RETURN > +L(return_vzeroupper): > + ZERO_UPPER_VEC_REGISTERS_RETURN > > # ifndef USE_AS_STRCHRNUL > L(zero): > @@ -103,7 +95,8 @@ L(zero): > > .p2align 4 > L(first_vec_x1): > - tzcntl %eax, %eax > + /* Use bsf to save code size. */ > + bsfl %eax, %eax > incq %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > @@ -113,9 +106,10 @@ L(first_vec_x1): > addq %rdi, %rax > VZEROUPPER_RETURN > > - .p2align 4 > + .p2align 4,, 10 > L(first_vec_x2): > - tzcntl %eax, %eax > + /* Use bsf to save code size. */ > + bsfl %eax, %eax > addq $(VEC_SIZE + 1), %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > @@ -125,9 +119,10 @@ L(first_vec_x2): > addq %rdi, %rax > VZEROUPPER_RETURN > > - .p2align 4 > + .p2align 4,, 8 > L(first_vec_x3): > - tzcntl %eax, %eax > + /* Use bsf to save code size. */ > + bsfl %eax, %eax > addq $(VEC_SIZE * 2 + 1), %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > @@ -137,6 +132,21 @@ L(first_vec_x3): > addq %rdi, %rax > VZEROUPPER_RETURN > > + .p2align 4,, 10 > +L(first_vec_x4): > + /* Use bsf to save code size. */ > + bsfl %eax, %eax > + addq $(VEC_SIZE * 3 + 1), %rdi > +# ifndef USE_AS_STRCHRNUL > + /* Found CHAR or the null byte. */ > + cmp (%rdi, %rax), %CHAR_REG > + jne L(zero) > +# endif > + addq %rdi, %rax > + VZEROUPPER_RETURN > + > + > + > .p2align 4 > L(aligned_more): > /* Align data to VEC_SIZE - 1. This is the same number of > @@ -146,90 +156,92 @@ L(aligned_more): > L(cross_page_continue): > /* Check the next 4 * VEC_SIZE. Only one VEC_SIZE at a time > since data is only aligned to VEC_SIZE. */ > - vmovdqa 1(%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqa 1(%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jnz L(first_vec_x1) > > - vmovdqa (VEC_SIZE + 1)(%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqa (VEC_SIZE + 1)(%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jnz L(first_vec_x2) > > - vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jnz L(first_vec_x3) > > - vmovdqa (VEC_SIZE * 3 + 1)(%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqa (VEC_SIZE * 3 + 1)(%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jnz L(first_vec_x4) > - /* Align data to VEC_SIZE * 4 - 1. */ > - addq $(VEC_SIZE * 4 + 1), %rdi > - andq $-(VEC_SIZE * 4), %rdi > + /* Align data to VEC_SIZE * 4 - 1. */ > + incq %rdi > + orq $(VEC_SIZE * 4 - 1), %rdi > .p2align 4 > L(loop_4x_vec): > /* Compare 4 * VEC at a time forward. */ > - vmovdqa (%rdi), %ymm5 > - vmovdqa (VEC_SIZE)(%rdi), %ymm6 > - vmovdqa (VEC_SIZE * 2)(%rdi), %ymm7 > - vmovdqa (VEC_SIZE * 3)(%rdi), %ymm8 > + vmovdqa 1(%rdi), %ymm6 > + vmovdqa (VEC_SIZE + 1)(%rdi), %ymm7 > > /* Leaves only CHARS matching esi as 0. */ > - vpxor %ymm5, %ymm0, %ymm1 > vpxor %ymm6, %ymm0, %ymm2 > vpxor %ymm7, %ymm0, %ymm3 > - vpxor %ymm8, %ymm0, %ymm4 > > - VPMINU %ymm1, %ymm5, %ymm1 > VPMINU %ymm2, %ymm6, %ymm2 > VPMINU %ymm3, %ymm7, %ymm3 > - VPMINU %ymm4, %ymm8, %ymm4 > > - VPMINU %ymm1, %ymm2, %ymm5 > - VPMINU %ymm3, %ymm4, %ymm6 > + vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm6 > + vmovdqa (VEC_SIZE * 3 + 1)(%rdi), %ymm7 > + > + vpxor %ymm6, %ymm0, %ymm4 > + vpxor %ymm7, %ymm0, %ymm5 > + > + VPMINU %ymm4, %ymm6, %ymm4 > + VPMINU %ymm5, %ymm7, %ymm5 > > - VPMINU %ymm5, %ymm6, %ymm6 > + VPMINU %ymm2, %ymm3, %ymm6 > + VPMINU %ymm4, %ymm5, %ymm7 > > - VPCMPEQ %ymm6, %ymm9, %ymm6 > - vpmovmskb %ymm6, %ecx > + VPMINU %ymm6, %ymm7, %ymm7 > + > + VPCMPEQ %ymm7, %ymm1, %ymm7 > + vpmovmskb %ymm7, %ecx > subq $-(VEC_SIZE * 4), %rdi > testl %ecx, %ecx > jz L(loop_4x_vec) > > - > - VPCMPEQ %ymm1, %ymm9, %ymm1 > - vpmovmskb %ymm1, %eax > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpmovmskb %ymm2, %eax > testl %eax, %eax > jnz L(last_vec_x0) > > > - VPCMPEQ %ymm5, %ymm9, %ymm2 > - vpmovmskb %ymm2, %eax > + VPCMPEQ %ymm3, %ymm1, %ymm3 > + vpmovmskb %ymm3, %eax > testl %eax, %eax > jnz L(last_vec_x1) > > - VPCMPEQ %ymm3, %ymm9, %ymm3 > - vpmovmskb %ymm3, %eax > + VPCMPEQ %ymm4, %ymm1, %ymm4 > + vpmovmskb %ymm4, %eax > /* rcx has combined result from all 4 VEC. It will only be used > if the first 3 other VEC all did not contain a match. */ > salq $32, %rcx > orq %rcx, %rax > tzcntq %rax, %rax > - subq $(VEC_SIZE * 2), %rdi > + subq $(VEC_SIZE * 2 - 1), %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > cmp (%rdi, %rax), %CHAR_REG > @@ -239,10 +251,11 @@ L(loop_4x_vec): > VZEROUPPER_RETURN > > > - .p2align 4 > + .p2align 4,, 10 > L(last_vec_x0): > - tzcntl %eax, %eax > - addq $-(VEC_SIZE * 4), %rdi > + /* Use bsf to save code size. */ > + bsfl %eax, %eax > + addq $-(VEC_SIZE * 4 - 1), %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > cmp (%rdi, %rax), %CHAR_REG > @@ -251,16 +264,11 @@ L(last_vec_x0): > addq %rdi, %rax > VZEROUPPER_RETURN > > -# ifndef USE_AS_STRCHRNUL > -L(zero_end): > - xorl %eax, %eax > - VZEROUPPER_RETURN > -# endif > > - .p2align 4 > + .p2align 4,, 10 > L(last_vec_x1): > tzcntl %eax, %eax > - subq $(VEC_SIZE * 3), %rdi > + subq $(VEC_SIZE * 3 - 1), %rdi > # ifndef USE_AS_STRCHRNUL > /* Found CHAR or the null byte. */ > cmp (%rdi, %rax), %CHAR_REG > @@ -269,18 +277,23 @@ L(last_vec_x1): > addq %rdi, %rax > VZEROUPPER_RETURN > > +# ifndef USE_AS_STRCHRNUL > +L(zero_end): > + xorl %eax, %eax > + VZEROUPPER_RETURN > +# endif > > /* Cold case for crossing page with first load. */ > - .p2align 4 > + .p2align 4,, 8 > L(cross_page_boundary): > movq %rdi, %rdx > /* Align rdi to VEC_SIZE - 1. */ > orq $(VEC_SIZE - 1), %rdi > - vmovdqa -(VEC_SIZE - 1)(%rdi), %ymm8 > - VPCMPEQ %ymm8, %ymm0, %ymm1 > - VPCMPEQ %ymm8, %ymm9, %ymm2 > - vpor %ymm1, %ymm2, %ymm1 > - vpmovmskb %ymm1, %eax > + vmovdqa -(VEC_SIZE - 1)(%rdi), %ymm2 > + VPCMPEQ %ymm2, %ymm0, %ymm3 > + VPCMPEQ %ymm2, %ymm1, %ymm2 > + vpor %ymm3, %ymm2, %ymm3 > + vpmovmskb %ymm3, %eax > /* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT > so no need to manually mod edx. */ > sarxl %edx, %eax, %eax > @@ -291,13 +304,10 @@ L(cross_page_boundary): > xorl %ecx, %ecx > /* Found CHAR or the null byte. */ > cmp (%rdx, %rax), %CHAR_REG > - leaq (%rdx, %rax), %rax > - cmovne %rcx, %rax > -# else > - addq %rdx, %rax > + jne L(zero_end) > # endif > -L(return_vzeroupper): > - ZERO_UPPER_VEC_REGISTERS_RETURN > + addq %rdx, %rax > + VZEROUPPER_RETURN > > END (STRCHR) > -# endif > +#endif > -- > 2.25.1 > -- H.J.