From patchwork Mon May 3 08:44:36 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Noah Goldstein X-Patchwork-Id: 43206 Return-Path: X-Original-To: patchwork@sourceware.org Delivered-To: patchwork@sourceware.org Received: from server2.sourceware.org (localhost [IPv6:::1]) by sourceware.org (Postfix) with ESMTP id 5E58D3939C34; Mon, 3 May 2021 08:45:55 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 5E58D3939C34 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1620031555; bh=V8uvesmtTTjDby+gldo8jGJpy7XFoft9+Ob3FMVGUMU=; h=To:Subject:Date:In-Reply-To:References:List-Id:List-Unsubscribe: List-Archive:List-Post:List-Help:List-Subscribe:From:Reply-To: From; b=YEmOHcm3G9B/bZ4agcl+hZWhaVJXeVU+P/QeAUTt76GnxYjRmIqsWJ2OBdKP5jhmX tNZu4LvKTd7/BfgBwBpqET9fn8hWEh6dPmkwbwnr9B7HMlmYJVh4giJpX6KGJ6oy96 jxKmuvXE9KvNKLsbOn+aRMUaYyszzZwGIeyP9V9s= X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-qt1-x832.google.com (mail-qt1-x832.google.com [IPv6:2607:f8b0:4864:20::832]) by sourceware.org (Postfix) with ESMTPS id CE22C385DC33 for ; Mon, 3 May 2021 08:45:51 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org CE22C385DC33 Received: by mail-qt1-x832.google.com with SMTP id f12so3128337qtf.2 for ; Mon, 03 May 2021 01:45:51 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:from:to:cc:subject:date:message-id:in-reply-to :references:mime-version:content-transfer-encoding; bh=V8uvesmtTTjDby+gldo8jGJpy7XFoft9+Ob3FMVGUMU=; b=eajGsdvZM0uNnfMytWdH4rsCf5dVONqyMc+Bs81aNB3/K4vuzft+GIC4BwJMwpDLkS bR+hWXV41QkjcoVpMU7GxIhanGdcxF/Tc5cx4TVX8wR+uSNUIXNS3Vy+Oi91uIroMGwn bkTK9AgKSYPAxTBhUMHYNB9ehbHLlsZZ834MYS/F7xOvefGmB93UGlLYNyOOWn0vgo69 WqZhrBo5zfxDIeRbM6uFOdhx8eIOhKn/GHgywCkLB8LTSckh+jQy+IbZNjsSxJvApSwF X/Vk0bWC5lJySgTv1vGiARV8HzbgNYIulaobDIRD3CQxid7j+lpgTW9m/kb7v3SjxS3l Gg3A== X-Gm-Message-State: AOAM530Uaz1rGjiBZWowD9b8U5Ticf9a/JXIGBjqTr4NcYBvpHN8w4ty JW3MC5WF6KZl+antq9YVKPdExs8T/6k= X-Google-Smtp-Source: ABdhPJw9SKCMXchXgZrIq4SFxnM8fqDU7F2zX4zX+mWn8ayKZWuM7ufHiTH4UOyfwKLJVBr+cs0O7Q== X-Received: by 2002:a05:622a:513:: with SMTP id l19mr8607036qtx.328.1620031550979; Mon, 03 May 2021 01:45:50 -0700 (PDT) Received: from localhost.localdomain (pool-71-245-178-39.pitbpa.fios.verizon.net. [71.245.178.39]) by smtp.googlemail.com with ESMTPSA id a3sm7954970qto.2.2021.05.03.01.45.50 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Mon, 03 May 2021 01:45:50 -0700 (PDT) To: libc-alpha@sourceware.org Subject: [PATCH v1 2/3] x86: Optimize memchr-avx2.S Date: Mon, 3 May 2021 04:44:36 -0400 Message-Id: <20210503084435.160548-2-goldstein.w.n@gmail.com> X-Mailer: git-send-email 2.29.2 In-Reply-To: <20210503084435.160548-1-goldstein.w.n@gmail.com> References: <20210503084435.160548-1-goldstein.w.n@gmail.com> MIME-Version: 1.0 X-Spam-Status: No, score=-12.6 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 autolearn=ham autolearn_force=no version=3.4.2 X-Spam-Checker-Version: SpamAssassin 3.4.2 (2018-09-13) 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-Patchwork-Original-From: Noah Goldstein via Libc-alpha From: Noah Goldstein Reply-To: Noah Goldstein Errors-To: libc-alpha-bounces@sourceware.org Sender: "Libc-alpha" No bug. This commit optimizes memchr-avx2.S. The optimizations include replacing some branches with cmovcc, avoiding some branches entirely in the less_4x_vec case, making the page cross logic less strict, asaving a few instructions the in loop return loop. test-memchr, test-rawmemchr, and test-wmemchr are all passing. Signed-off-by: Noah Goldstein --- sysdeps/x86_64/multiarch/memchr-avx2.S | 446 +++++++++++++++---------- 1 file changed, 262 insertions(+), 184 deletions(-) diff --git a/sysdeps/x86_64/multiarch/memchr-avx2.S b/sysdeps/x86_64/multiarch/memchr-avx2.S index 1fcb1c350f..8368fcd1e1 100644 --- a/sysdeps/x86_64/multiarch/memchr-avx2.S +++ b/sysdeps/x86_64/multiarch/memchr-avx2.S @@ -26,8 +26,22 @@ # ifdef USE_AS_WMEMCHR # define VPCMPEQ vpcmpeqd +# define VPBROADCAST vpbroadcastd +# define CHAR_SIZE 4 # else # define VPCMPEQ vpcmpeqb +# define VPBROADCAST vpbroadcastb +# define CHAR_SIZE 1 +# endif + +# ifdef USE_AS_RAWMEMCHR +# define ERAW_PTR_REG ecx +# define RRAW_PTR_REG rcx +# define ALGN_PTR_REG rdi +# else +# define ERAW_PTR_REG edi +# define RRAW_PTR_REG rdi +# define ALGN_PTR_REG rcx # endif # ifndef VZEROUPPER @@ -39,303 +53,367 @@ # endif # define VEC_SIZE 32 +# define PAGE_SIZE 4096 + .section SECTION(.text),"ax",@progbits -ENTRY (MEMCHR) +ENTRY(MEMCHR) # ifndef USE_AS_RAWMEMCHR /* Check for zero length. */ test %RDX_LP, %RDX_LP jz L(null) # endif - movl %edi, %ecx - /* Broadcast CHAR to YMM0. */ - vmovd %esi, %xmm0 # ifdef USE_AS_WMEMCHR shl $2, %RDX_LP - vpbroadcastd %xmm0, %ymm0 # else # ifdef __ILP32__ /* Clear the upper 32 bits. */ movl %edx, %edx # endif - vpbroadcastb %xmm0, %ymm0 # endif - /* Check if we may cross page boundary with one vector load. */ - andl $(2 * VEC_SIZE - 1), %ecx - cmpl $VEC_SIZE, %ecx - ja L(cros_page_boundary) + /* Broadcast CHAR to YMMMATCH. */ + vmovd %esi, %xmm0 + VPBROADCAST %xmm0, %ymm0 + /* Check if we may cross page boundary with one + vector load. */ + movl %edi, %eax + andl $(PAGE_SIZE - 1), %eax + cmpl $(PAGE_SIZE - VEC_SIZE), %eax + ja L(cross_page_boundary) /* Check the first VEC_SIZE bytes. */ - VPCMPEQ (%rdi), %ymm0, %ymm1 + VPCMPEQ (%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax - testl %eax, %eax - # ifndef USE_AS_RAWMEMCHR - jnz L(first_vec_x0_check) - /* Adjust length and check the end of data. */ - subq $VEC_SIZE, %rdx - jbe L(zero) -# else - jnz L(first_vec_x0) + /* If length < CHAR_PER_VEC handle special. */ + cmpq $VEC_SIZE, %rdx + jbe L(first_vec_x0) # endif - - /* Align data for aligned loads in the loop. */ - addq $VEC_SIZE, %rdi - andl $(VEC_SIZE - 1), %ecx - andq $-VEC_SIZE, %rdi + testl %eax, %eax + jz L(aligned_more) + tzcntl %eax, %eax + addq %rdi, %rax + VZEROUPPER_RETURN # ifndef USE_AS_RAWMEMCHR - /* Adjust length. */ - addq %rcx, %rdx - - subq $(VEC_SIZE * 4), %rdx - jbe L(last_4x_vec_or_less) + .p2align 5 +L(first_vec_x0): + /* Check if first match was before length. */ + tzcntl %eax, %eax + xorl %ecx, %ecx + cmpl %eax, %edx + leaq (%rdi, %rax), %rax + cmovle %rcx, %rax + VZEROUPPER_RETURN +L(null): + xorl %eax, %eax + ret # endif - jmp L(more_4x_vec) - .p2align 4 -L(cros_page_boundary): - andl $(VEC_SIZE - 1), %ecx - andq $-VEC_SIZE, %rdi - VPCMPEQ (%rdi), %ymm0, %ymm1 +L(cross_page_boundary): + /* Save pointer before aligning as its original + value is necessary for computer return address if byte is + found or adjusting length if it is not and this is + memchr. */ + movq %rdi, %rcx + /* Align data to VEC_SIZE - 1. ALGN_PTR_REG is + rcx for memchr and rdi for rawmemchr. */ + orq $(VEC_SIZE - 1), %ALGN_PTR_REG + VPCMPEQ -(VEC_SIZE - 1)(%ALGN_PTR_REG), %ymm0, %ymm1 vpmovmskb %ymm1, %eax +# ifndef USE_AS_RAWMEMCHR + /* Calculate length until end of page (length + checked for a match). */ + leaq 1(%ALGN_PTR_REG), %rsi + subq %RRAW_PTR_REG, %rsi +# endif /* Remove the leading bytes. */ - sarl %cl, %eax - testl %eax, %eax - jz L(aligned_more) - tzcntl %eax, %eax + sarxl %ERAW_PTR_REG, %eax, %eax # ifndef USE_AS_RAWMEMCHR /* Check the end of data. */ - cmpq %rax, %rdx - jbe L(zero) + cmpq %rsi, %rdx + jbe L(first_vec_x0) # endif - addq %rdi, %rax - addq %rcx, %rax + testl %eax, %eax + jz L(cross_page_continue) + tzcntl %eax, %eax + addq %RRAW_PTR_REG, %rax L(return_vzeroupper): ZERO_UPPER_VEC_REGISTERS_RETURN .p2align 4 -L(aligned_more): -# ifndef USE_AS_RAWMEMCHR - /* Calculate "rdx + rcx - VEC_SIZE" with "rdx - (VEC_SIZE - rcx)" - instead of "(rdx + rcx) - VEC_SIZE" to void possible addition - overflow. */ - negq %rcx - addq $VEC_SIZE, %rcx +L(first_vec_x1): + tzcntl %eax, %eax + incq %rdi + addq %rdi, %rax + VZEROUPPER_RETURN - /* Check the end of data. */ - subq %rcx, %rdx - jbe L(zero) -# endif + .p2align 4 +L(first_vec_x2): + tzcntl %eax, %eax + addq $(VEC_SIZE + 1), %rdi + addq %rdi, %rax + VZEROUPPER_RETURN - addq $VEC_SIZE, %rdi + .p2align 4 +L(first_vec_x3): + tzcntl %eax, %eax + addq $(VEC_SIZE * 2 + 1), %rdi + addq %rdi, %rax + VZEROUPPER_RETURN -# ifndef USE_AS_RAWMEMCHR - subq $(VEC_SIZE * 4), %rdx - jbe L(last_4x_vec_or_less) -# endif -L(more_4x_vec): - /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time - since data is only aligned to VEC_SIZE. */ - VPCMPEQ (%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x0) + .p2align 4 +L(first_vec_x4): + tzcntl %eax, %eax + addq $(VEC_SIZE * 3 + 1), %rdi + addq %rdi, %rax + VZEROUPPER_RETURN - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 + .p2align 4 +L(aligned_more): + /* Check the first 4 * VEC_SIZE. Only one + VEC_SIZE at a time since data is only aligned to + VEC_SIZE. */ + +# ifndef USE_AS_RAWMEMCHR +L(cross_page_continue): + /* Align data to VEC_SIZE - 1. */ + xorl %ecx, %ecx + subl %edi, %ecx + orq $(VEC_SIZE - 1), %rdi + /* esi is for adjusting length to see if near the + end. */ + leal (VEC_SIZE * 4 + 1)(%rdi, %rcx), %esi +# else + orq $(VEC_SIZE - 1), %rdi +L(cross_page_continue): +# endif + /* Load first VEC regardless. */ + VPCMPEQ 1(%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax +# ifndef USE_AS_RAWMEMCHR + /* Adjust length. If near end handle specially. + */ + subq %rsi, %rdx + jbe L(last_4x_vec_or_less) +# endif testl %eax, %eax jnz L(first_vec_x1) - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1 + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax testl %eax, %eax jnz L(first_vec_x2) - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1 + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax testl %eax, %eax jnz L(first_vec_x3) - addq $(VEC_SIZE * 4), %rdi + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + testl %eax, %eax + jnz L(first_vec_x4) # ifndef USE_AS_RAWMEMCHR + /* Check if at last VEC_SIZE * 4 length. */ subq $(VEC_SIZE * 4), %rdx - jbe L(last_4x_vec_or_less) -# endif - - /* Align data to 4 * VEC_SIZE. */ - movq %rdi, %rcx - andl $(4 * VEC_SIZE - 1), %ecx - andq $-(4 * VEC_SIZE), %rdi - -# ifndef USE_AS_RAWMEMCHR - /* Adjust length. */ + jbe L(last_4x_vec_or_less_cmpeq) + /* Align data to VEC_SIZE * 4 - 1 for the loop + and readjust length. */ + incq %rdi + movl %edi, %ecx + orq $(VEC_SIZE * 4 - 1), %rdi + andl $(VEC_SIZE * 4 - 1), %ecx addq %rcx, %rdx +# else + /* Align data to VEC_SIZE * 4 - 1 for loop. */ + incq %rdi + orq $(VEC_SIZE * 4 - 1), %rdi # endif + /* Compare 4 * VEC at a time forward. */ .p2align 4 L(loop_4x_vec): - /* Compare 4 * VEC at a time forward. */ - VPCMPEQ (%rdi), %ymm0, %ymm1 - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm2 - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm3 - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm4 - + VPCMPEQ 1(%rdi), %ymm0, %ymm1 + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm2 + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm3 + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm4 vpor %ymm1, %ymm2, %ymm5 vpor %ymm3, %ymm4, %ymm6 vpor %ymm5, %ymm6, %ymm5 - vpmovmskb %ymm5, %eax - testl %eax, %eax - jnz L(4x_vec_end) - - addq $(VEC_SIZE * 4), %rdi - + vpmovmskb %ymm5, %ecx # ifdef USE_AS_RAWMEMCHR - jmp L(loop_4x_vec) + subq $-(VEC_SIZE * 4), %rdi + testl %ecx, %ecx + jz L(loop_4x_vec) # else - subq $(VEC_SIZE * 4), %rdx - ja L(loop_4x_vec) + testl %ecx, %ecx + jnz L(loop_4x_vec_end) -L(last_4x_vec_or_less): - /* Less than 4 * VEC and aligned to VEC_SIZE. */ - addl $(VEC_SIZE * 2), %edx - jle L(last_2x_vec) + subq $-(VEC_SIZE * 4), %rdi - VPCMPEQ (%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x0) + subq $(VEC_SIZE * 4), %rdx + ja L(loop_4x_vec) - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 + /* Fall through into less than 4 remaining + vectors of length case. */ + VPCMPEQ (VEC_SIZE * 0 + 1)(%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax + .p2align 4 +L(last_4x_vec_or_less): + /* Check if first VEC contained match. */ testl %eax, %eax - jnz L(first_vec_x1) + jnz L(first_vec_x1_check) - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax + /* If remaining length > VEC_SIZE * 2. */ + addl $(VEC_SIZE * 2), %edx + jg L(last_4x_vec) - jnz L(first_vec_x2_check) - subl $VEC_SIZE, %edx - jle L(zero) +L(last_2x_vec): + /* If remaining length < VEC_SIZE. */ + addl $VEC_SIZE, %edx + jle L(zero_end) - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1 + /* Check VEC2 and compare any match with + remaining length. */ + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1 vpmovmskb %ymm1, %eax - testl %eax, %eax - - jnz L(first_vec_x3_check) - xorl %eax, %eax + tzcntl %eax, %eax + cmpl %eax, %edx + jbe L(set_zero_end) + addq $(VEC_SIZE + 1), %rdi + addq %rdi, %rax +L(zero_end): VZEROUPPER_RETURN .p2align 4 -L(last_2x_vec): - addl $(VEC_SIZE * 2), %edx - VPCMPEQ (%rdi), %ymm0, %ymm1 +L(loop_4x_vec_end): +# endif + /* rawmemchr will fall through into this if match + was found in loop. */ + vpmovmskb %ymm1, %eax testl %eax, %eax + jnz L(last_vec_x1_return) - jnz L(first_vec_x0_check) - subl $VEC_SIZE, %edx - jle L(zero) - - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax + vpmovmskb %ymm2, %eax testl %eax, %eax - jnz L(first_vec_x1_check) - xorl %eax, %eax - VZEROUPPER_RETURN + jnz L(last_vec_x2_return) - .p2align 4 -L(first_vec_x0_check): - tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rdx - jbe L(zero) + vpmovmskb %ymm3, %eax + /* Combine VEC3 matches (eax) with VEC4 matches + (ecx). */ + salq $32, %rcx + orq %rcx, %rax + tzcntq %rax, %rax +# ifdef USE_AS_RAWMEMCHR + subq $(VEC_SIZE * 2 - 1), %rdi +# else + subq $-(VEC_SIZE * 2 + 1), %rdi +# endif addq %rdi, %rax VZEROUPPER_RETURN +# ifndef USE_AS_RAWMEMCHR .p2align 4 L(first_vec_x1_check): tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rdx - jbe L(zero) - addq $VEC_SIZE, %rax + /* Adjust length. */ + subl $-(VEC_SIZE * 4), %edx + /* Check if match within remaining length. */ + cmpl %eax, %edx + jbe L(set_zero_end) + incq %rdi addq %rdi, %rax VZEROUPPER_RETURN + .p2align 4 +L(set_zero_end): + xorl %eax, %eax + VZEROUPPER_RETURN +# endif .p2align 4 -L(first_vec_x2_check): +L(last_vec_x1_return): tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rdx - jbe L(zero) - addq $(VEC_SIZE * 2), %rax +# ifdef USE_AS_RAWMEMCHR + subq $(VEC_SIZE * 4 - 1), %rdi +# else + incq %rdi +# endif addq %rdi, %rax VZEROUPPER_RETURN .p2align 4 -L(first_vec_x3_check): +L(last_vec_x2_return): tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rdx - jbe L(zero) - addq $(VEC_SIZE * 3), %rax +# ifdef USE_AS_RAWMEMCHR + subq $(VEC_SIZE * 3 - 1), %rdi +# else + subq $-(VEC_SIZE + 1), %rdi +# endif addq %rdi, %rax VZEROUPPER_RETURN +# ifndef USE_AS_RAWMEMCHR .p2align 4 -L(zero): - xorl %eax, %eax - jmp L(return_vzeroupper) +L(last_4x_vec_or_less_cmpeq): + VPCMPEQ (VEC_SIZE * 4 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + subq $-(VEC_SIZE * 4), %rdi + /* Check first VEC regardless. */ + testl %eax, %eax + jnz L(first_vec_x1_check) + /* If remaining length <= CHAR_PER_VEC * 2. */ + addl $(VEC_SIZE * 2), %edx + jle L(last_2x_vec) .p2align 4 -L(null): - xorl %eax, %eax - ret -# endif +L(last_4x_vec): + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + testl %eax, %eax + jnz L(last_vec_x2_return) - .p2align 4 -L(first_vec_x0): - tzcntl %eax, %eax - addq %rdi, %rax - VZEROUPPER_RETURN + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax - .p2align 4 -L(first_vec_x1): - tzcntl %eax, %eax - addq $VEC_SIZE, %rax - addq %rdi, %rax - VZEROUPPER_RETURN + /* Create mask for possible matches within + remaining length. */ + movq $-1, %rcx + bzhiq %rdx, %rcx, %rcx - .p2align 4 -L(first_vec_x2): + /* Test matches in data against length match. */ + andl %ecx, %eax + jnz L(last_vec_x3) + + /* if remaining length <= VEC_SIZE * 3 (Note this + is after remaining length was found to be > VEC_SIZE * 2. + */ + subl $VEC_SIZE, %edx + jbe L(zero_end2) + + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + /* Shift remaining length mask for last VEC. */ + shrq $32, %rcx + andl %ecx, %eax + jz L(zero_end2) tzcntl %eax, %eax - addq $(VEC_SIZE * 2), %rax + addq $(VEC_SIZE * 3 + 1), %rdi addq %rdi, %rax +L(zero_end2): VZEROUPPER_RETURN .p2align 4 -L(4x_vec_end): - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x0) - vpmovmskb %ymm2, %eax - testl %eax, %eax - jnz L(first_vec_x1) - vpmovmskb %ymm3, %eax - testl %eax, %eax - jnz L(first_vec_x2) - vpmovmskb %ymm4, %eax - testl %eax, %eax -L(first_vec_x3): +L(last_vec_x3): tzcntl %eax, %eax - addq $(VEC_SIZE * 3), %rax + subq $-(VEC_SIZE * 2 + 1), %rdi addq %rdi, %rax VZEROUPPER_RETURN +# endif -END (MEMCHR) +END(MEMCHR) #endif