From patchwork Mon Apr 19 16:28:58 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Noah Goldstein X-Patchwork-Id: 43043 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 34D203956837; Mon, 19 Apr 2021 16:29:20 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 34D203956837 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1618849760; bh=JuJf0JMLHJs9pZ3+PDBAxep/vOMs238+iA9hg/eNMyQ=; h=To:Subject:Date:List-Id:List-Unsubscribe:List-Archive:List-Post: List-Help:List-Subscribe:From:Reply-To:From; b=aHTxi4KSBpeImTy5HrdprznKLTXxpTbdTq35cKHM6moSPK5kGO9FkMeSE85ApsG/v 56zYw2fwQXNyNczxmqQjtDlzkixv1KslWDwijsVB+Q+Go9JIIsDgPalfl0UU4DPA8h GSEdYVpY43Yu73PiQPEpFnfg9/lLvRFJLmi88iKk= X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-qk1-x72c.google.com (mail-qk1-x72c.google.com [IPv6:2607:f8b0:4864:20::72c]) by sourceware.org (Postfix) with ESMTPS id F37CB395182D for ; Mon, 19 Apr 2021 16:29:08 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org F37CB395182D Received: by mail-qk1-x72c.google.com with SMTP id q136so15419348qka.7 for ; Mon, 19 Apr 2021 09:29:08 -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:mime-version :content-transfer-encoding; bh=JuJf0JMLHJs9pZ3+PDBAxep/vOMs238+iA9hg/eNMyQ=; b=CpV8jxp9LfpgeZAOZRQO4BlC/b7CZxd7x3Qi0Ymos7CnrUSMVBGSeoCd4+JDsU5A+Y JkIah7TFDG4oh3ee/+koWPQZmHOu7gMt6Z3oC8z2d6q2kfAEhsvXkdmQ57aEwyGiYgr+ 6jnrroETeG7ewYIvONzhxVDVpQT6VbP7rH4Bi5N2fT2xvdQ/E8A4xjrxcCgpOmUhbjKr jw+ieW/qbvOdZDJaDb5MIgjaMNc6ZvfP8rM9G/rtlTWeQmSkg1Ul/Mzbjdwfq8V0qiJL OxOEVZr/EnHLKBXI4CRHVh/FOqFDWn1ygcsl3arrnW8KTHBnsKjllrhczRsULey/ceFf BW3w== X-Gm-Message-State: AOAM532VsSXAK8l5CyDNQivEGPxvew2Os11cQTIQeFLeu9NdCZ4ii3LL M0KEz3wDVVpSMcFuBmgyFDhY02VS7ZU= X-Google-Smtp-Source: ABdhPJxmhuI2PubnVJbkBcEyJ2XAZafYdpHrV55Mh+8aq1cM4njnjLs8Ie3J3FCc0HSLGsMWhTQNCw== X-Received: by 2002:a05:620a:d5e:: with SMTP id o30mr12568266qkl.288.1618849746775; Mon, 19 Apr 2021 09:29:06 -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 c23sm9573332qtm.46.2021.04.19.09.29.05 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Mon, 19 Apr 2021 09:29:06 -0700 (PDT) To: libc-alpha@sourceware.org Subject: [PATCH v3 1/2] x86: Optimize strlen-evex.S Date: Mon, 19 Apr 2021 12:28:58 -0400 Message-Id: <20210419162859.2285431-1-goldstein.w.n@gmail.com> X-Mailer: git-send-email 2.29.2 MIME-Version: 1.0 X-Spam-Status: No, score=-12.5 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 strlen-evex.S. The optimizations are mostly small things but they add up to roughly 10-30% performance improvement for strlen. The results for strnlen are bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen are all passing. Signed-off-by: Noah Goldstein --- sysdeps/x86_64/multiarch/strlen-evex.S | 584 ++++++++++++++----------- 1 file changed, 320 insertions(+), 264 deletions(-) diff --git a/sysdeps/x86_64/multiarch/strlen-evex.S b/sysdeps/x86_64/multiarch/strlen-evex.S index 0583819078..36b28198c8 100644 --- a/sysdeps/x86_64/multiarch/strlen-evex.S +++ b/sysdeps/x86_64/multiarch/strlen-evex.S @@ -29,11 +29,13 @@ # ifdef USE_AS_WCSLEN # define VPCMP vpcmpd # define VPMINU vpminud -# define SHIFT_REG r9d +# define SHIFT_REG ecx +# define CHAR_SIZE 4 # else # define VPCMP vpcmpb # define VPMINU vpminub -# define SHIFT_REG ecx +# define SHIFT_REG edx +# define CHAR_SIZE 1 # endif # define XMMZERO xmm16 @@ -46,132 +48,168 @@ # define YMM6 ymm22 # define VEC_SIZE 32 +# define PAGE_SIZE 4096 +# define LOG_PAGE_SIZE 12 +# define CHAR_PER_VEC (VEC_SIZE / CHAR_SIZE) .section .text.evex,"ax",@progbits ENTRY (STRLEN) # ifdef USE_AS_STRNLEN - /* Check for zero length. */ + /* Check zero length. */ test %RSI_LP, %RSI_LP jz L(zero) -# ifdef USE_AS_WCSLEN - shl $2, %RSI_LP -# elif defined __ILP32__ +# ifdef __ILP32__ /* Clear the upper 32 bits. */ movl %esi, %esi # endif mov %RSI_LP, %R8_LP # endif - movl %edi, %ecx - movq %rdi, %rdx + movl %edi, %eax vpxorq %XMMZERO, %XMMZERO, %XMMZERO - + /* Shift left eax to clear all bits not relevant to page cross + check. This saves 2 bytes of code as opposed to using andl with + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted + left by the same amount (an imm32 either way). */ + sall $(32 - LOG_PAGE_SIZE), %eax /* 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) + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax + ja L(cross_page_boundary) /* Check the first VEC_SIZE bytes. Each bit in K0 represents a null byte. */ VPCMP $0, (%rdi), %YMMZERO, %k0 kmovd %k0, %eax - testl %eax, %eax - # ifdef USE_AS_STRNLEN - jnz L(first_vec_x0_check) - /* Adjust length and check the end of data. */ - subq $VEC_SIZE, %rsi - jbe L(max) -# else - jnz L(first_vec_x0) + /* If length < CHAR_PER_VEC handle special. */ + cmpq $CHAR_PER_VEC, %rsi + 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 + ret # ifdef USE_AS_STRNLEN - /* Adjust length. */ - addq %rcx, %rsi +L(zero): + xorl %eax, %eax + ret - subq $(VEC_SIZE * 4), %rsi - jbe L(last_4x_vec_or_less) + .p2align 4 +L(first_vec_x0): + /* Set bit for max len so that tzcnt will return min of max len + and position of first match. */ + btsq %rsi, %rax + tzcntl %eax, %eax + ret # endif - jmp L(more_4x_vec) .p2align 4 -L(cros_page_boundary): - andl $(VEC_SIZE - 1), %ecx - andq $-VEC_SIZE, %rdi - -# ifdef USE_AS_WCSLEN - /* NB: Divide shift count by 4 since each bit in K0 represent 4 - bytes. */ - movl %ecx, %SHIFT_REG - sarl $2, %SHIFT_REG +L(first_vec_x1): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ +# ifdef USE_AS_STRNLEN + /* Use ecx which was computed earlier to compute correct value. + */ + leal -(CHAR_PER_VEC * 4 + 1)(%rcx, %rax), %eax +# else + subl %edx, %edi +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %edi +# endif + leal CHAR_PER_VEC(%rdi, %rax), %eax # endif - VPCMP $0, (%rdi), %YMMZERO, %k0 - kmovd %k0, %eax + ret - /* Remove the leading bytes. */ - sarxl %SHIFT_REG, %eax, %eax - testl %eax, %eax - jz L(aligned_more) + .p2align 4 +L(first_vec_x2): tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) -# endif - addq %rdi, %rax - addq %rcx, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax + /* Use ecx which was computed earlier to compute correct value. + */ + leal -(CHAR_PER_VEC * 3 + 1)(%rcx, %rax), %eax +# else + subl %edx, %edi +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %edi +# endif + leal (CHAR_PER_VEC * 2)(%rdi, %rax), %eax # endif ret .p2align 4 -L(aligned_more): +L(first_vec_x3): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - /* "rcx" is less than VEC_SIZE. 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 - - /* Check the end of data. */ - subq %rcx, %rsi - jbe L(max) + /* Use ecx which was computed earlier to compute correct value. + */ + leal -(CHAR_PER_VEC * 2 + 1)(%rcx, %rax), %eax +# else + subl %edx, %edi +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %edi +# endif + leal (CHAR_PER_VEC * 3)(%rdi, %rax), %eax # endif + ret - addq $VEC_SIZE, %rdi - + .p2align 4 +L(first_vec_x4): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - subq $(VEC_SIZE * 4), %rsi - jbe L(last_4x_vec_or_less) + /* Use ecx which was computed earlier to compute correct value. + */ + leal -(CHAR_PER_VEC + 1)(%rcx, %rax), %eax +# else + subl %edx, %edi +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %edi +# endif + leal (CHAR_PER_VEC * 4)(%rdi, %rax), %eax # endif + ret -L(more_4x_vec): + .p2align 5 +L(aligned_more): + movq %rdi, %rdx + /* Align data to VEC_SIZE. */ + andq $-(VEC_SIZE), %rdi +L(cross_page_continue): /* Check the first 4 * VEC_SIZE. Only one VEC_SIZE at a time since data is only aligned to VEC_SIZE. */ - VPCMP $0, (%rdi), %YMMZERO, %k0 - kmovd %k0, %eax - testl %eax, %eax - jnz L(first_vec_x0) - +# ifdef USE_AS_STRNLEN + /* + CHAR_SIZE because it simplies the logic in + last_4x_vec_or_less. */ + leaq (VEC_SIZE * 5 + CHAR_SIZE)(%rdi), %rcx + subq %rdx, %rcx +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %ecx +# endif +# endif + /* Load first VEC regardless. */ VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0 +# ifdef USE_AS_STRNLEN + /* Adjust length. If near end handle specially. */ + subq %rcx, %rsi + jb L(last_4x_vec_or_less) +# endif kmovd %k0, %eax testl %eax, %eax jnz L(first_vec_x1) VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0 kmovd %k0, %eax - testl %eax, %eax + test %eax, %eax jnz L(first_vec_x2) VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0 @@ -179,258 +217,276 @@ L(more_4x_vec): testl %eax, %eax jnz L(first_vec_x3) - addq $(VEC_SIZE * 4), %rdi - -# ifdef USE_AS_STRNLEN - subq $(VEC_SIZE * 4), %rsi - 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 + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0 + kmovd %k0, %eax + testl %eax, %eax + jnz L(first_vec_x4) + addq $VEC_SIZE, %rdi # ifdef USE_AS_STRNLEN - /* Adjust length. */ + /* Check if at last VEC_SIZE * 4 length. */ + cmpq $(CHAR_PER_VEC * 4 - 1), %rsi + jbe L(last_4x_vec_or_less_load) + movl %edi, %ecx + andl $(VEC_SIZE * 4 - 1), %ecx +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarl $2, %ecx +# endif + /* Readjust length. */ addq %rcx, %rsi # endif + /* Align data to VEC_SIZE * 4. */ + andq $-(VEC_SIZE * 4), %rdi + /* Compare 4 * VEC at a time forward. */ .p2align 4 L(loop_4x_vec): - /* Compare 4 * VEC at a time forward. */ - VMOVA (%rdi), %YMM1 - VMOVA VEC_SIZE(%rdi), %YMM2 - VMOVA (VEC_SIZE * 2)(%rdi), %YMM3 - VMOVA (VEC_SIZE * 3)(%rdi), %YMM4 - - VPMINU %YMM1, %YMM2, %YMM5 - VPMINU %YMM3, %YMM4, %YMM6 + /* Load first VEC regardless. */ + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1 +# ifdef USE_AS_STRNLEN + /* Break if at end of length. */ + subq $(CHAR_PER_VEC * 4), %rsi + jb L(last_4x_vec_or_less_cmpeq) +# endif + /* Save some code size by microfusing VPMINU with the load. Since + the matches in ymm2/ymm4 can only be returned if there where no + matches in ymm1/ymm3 respectively there is no issue with overlap. + */ + VPMINU (VEC_SIZE * 5)(%rdi), %YMM1, %YMM2 + VMOVA (VEC_SIZE * 6)(%rdi), %YMM3 + VPMINU (VEC_SIZE * 7)(%rdi), %YMM3, %YMM4 + + VPCMP $0, %YMM2, %YMMZERO, %k0 + VPCMP $0, %YMM4, %YMMZERO, %k1 + subq $-(VEC_SIZE * 4), %rdi + kortestd %k0, %k1 + jz L(loop_4x_vec) + + /* Check if end was in first half. */ + kmovd %k0, %eax + subq %rdx, %rdi +# ifdef USE_AS_WCSLEN + shrq $2, %rdi +# endif + testl %eax, %eax + jz L(second_vec_return) - VPMINU %YMM5, %YMM6, %YMM5 - VPCMP $0, %YMM5, %YMMZERO, %k0 - ktestd %k0, %k0 - jnz L(4x_vec_end) + VPCMP $0, %YMM1, %YMMZERO, %k2 + kmovd %k2, %edx + /* Combine VEC1 matches (edx) with VEC2 matches (eax). */ +# ifdef USE_AS_WCSLEN + sall $CHAR_PER_VEC, %eax + orl %edx, %eax + tzcntl %eax, %eax +# else + salq $CHAR_PER_VEC, %rax + orq %rdx, %rax + tzcntq %rax, %rax +# endif + addq %rdi, %rax + ret - addq $(VEC_SIZE * 4), %rdi -# ifndef USE_AS_STRNLEN - jmp L(loop_4x_vec) -# else - subq $(VEC_SIZE * 4), %rsi - ja L(loop_4x_vec) +# ifdef USE_AS_STRNLEN +L(last_4x_vec_or_less_load): + /* Depending on entry adjust rdi / prepare first VEC in YMM1. */ + VMOVA (VEC_SIZE * 4)(%rdi), %YMM1 +L(last_4x_vec_or_less_cmpeq): + VPCMP $0, %YMM1, %YMMZERO, %k0 + addq $(VEC_SIZE * 3), %rdi L(last_4x_vec_or_less): - /* Less than 4 * VEC and aligned to VEC_SIZE. */ - addl $(VEC_SIZE * 2), %esi - jle L(last_2x_vec) - - VPCMP $0, (%rdi), %YMMZERO, %k0 kmovd %k0, %eax + /* If remaining length > VEC_SIZE * 2. This works if esi is off by + VEC_SIZE * 4. */ + testl $(CHAR_PER_VEC * 2), %esi + jnz L(last_4x_vec) + + /* length may have been negative or positive by an offset of + CHAR_PER_VEC * 4 depending on where this was called from. This + fixes that. */ + andl $(CHAR_PER_VEC * 4 - 1), %esi testl %eax, %eax - jnz L(first_vec_x0) + jnz L(last_vec_x1_check) - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0 - kmovd %k0, %eax - testl %eax, %eax - jnz L(first_vec_x1) + /* Check the end of data. */ + subl $CHAR_PER_VEC, %esi + jb L(max) VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0 kmovd %k0, %eax - testl %eax, %eax - jnz L(first_vec_x2_check) - subl $VEC_SIZE, %esi - jle L(max) + tzcntl %eax, %eax + /* Check the end of data. */ + cmpl %eax, %esi + jb L(max) - VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0 - kmovd %k0, %eax - testl %eax, %eax - jnz L(first_vec_x3_check) + subq %rdx, %rdi +# ifdef USE_AS_WCSLEN + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi +# endif + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax + ret +L(max): movq %r8, %rax + ret +# endif + + /* Placed here in strnlen so that the jcc L(last_4x_vec_or_less) + in the 4x VEC loop can use 2 byte encoding. */ + .p2align 4 +L(second_vec_return): + VPCMP $0, %YMM3, %YMMZERO, %k0 + /* Combine YMM3 matches (k0) with YMM4 matches (k1). */ +# ifdef USE_AS_WCSLEN + kunpckbw %k0, %k1, %k0 + kmovd %k0, %eax + tzcntl %eax, %eax +# else + kunpckdq %k0, %k1, %k0 + kmovq %k0, %rax + tzcntq %rax, %rax +# endif + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax + ret + + +# ifdef USE_AS_STRNLEN +L(last_vec_x1_check): + tzcntl %eax, %eax + /* Check the end of data. */ + cmpl %eax, %esi + jb L(max) + subq %rdx, %rdi # ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi # endif + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax ret .p2align 4 -L(last_2x_vec): - addl $(VEC_SIZE * 2), %esi +L(last_4x_vec): + /* Test first 2x VEC normally. */ + testl %eax, %eax + jnz L(last_vec_x1) - VPCMP $0, (%rdi), %YMMZERO, %k0 + VPCMP $0, (VEC_SIZE * 2)(%rdi), %YMMZERO, %k0 kmovd %k0, %eax testl %eax, %eax - jnz L(first_vec_x0_check) - subl $VEC_SIZE, %esi - jle L(max) + jnz L(last_vec_x2) - VPCMP $0, VEC_SIZE(%rdi), %YMMZERO, %k0 + /* Normalize length. */ + andl $(CHAR_PER_VEC * 4 - 1), %esi + VPCMP $0, (VEC_SIZE * 3)(%rdi), %YMMZERO, %k0 kmovd %k0, %eax testl %eax, %eax - jnz L(first_vec_x1_check) - movq %r8, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif - ret + jnz L(last_vec_x3) - .p2align 4 -L(first_vec_x0_check): + /* Check the end of data. */ + subl $(CHAR_PER_VEC * 3), %esi + jb L(max) + + VPCMP $0, (VEC_SIZE * 4)(%rdi), %YMMZERO, %k0 + kmovd %k0, %eax tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq %rdi, %rax - subq %rdx, %rax + cmpl %eax, %esi + jb L(max_end) + + subq %rdx, %rdi # ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi # endif + leaq (CHAR_PER_VEC * 4)(%rdi, %rax), %rax ret .p2align 4 -L(first_vec_x1_check): +L(last_vec_x1): tzcntl %eax, %eax + subq %rdx, %rdi # ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $VEC_SIZE, %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi # endif + leaq (CHAR_PER_VEC)(%rdi, %rax), %rax ret .p2align 4 -L(first_vec_x2_check): +L(last_vec_x2): tzcntl %eax, %eax + subq %rdx, %rdi # ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $(VEC_SIZE * 2), %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi # endif + leaq (CHAR_PER_VEC * 2)(%rdi, %rax), %rax ret .p2align 4 -L(first_vec_x3_check): +L(last_vec_x3): tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif + subl $(CHAR_PER_VEC * 2), %esi /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $(VEC_SIZE * 3), %rax - addq %rdi, %rax - subq %rdx, %rax + cmpl %eax, %esi + jb L(max_end) + subq %rdx, %rdi # ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide bytes by 4 to get the wchar_t count. */ + sarq $2, %rdi # endif + leaq (CHAR_PER_VEC * 3)(%rdi, %rax), %rax ret - - .p2align 4 -L(max): +L(max_end): movq %r8, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif - ret - - .p2align 4 -L(zero): - xorl %eax, %eax ret # endif + /* Cold case for crossing page with first load. */ .p2align 4 -L(first_vec_x0): - tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - addq %rdi, %rax - subq %rdx, %rax +L(cross_page_boundary): + movq %rdi, %rdx + /* Align data to VEC_SIZE. */ + andq $-VEC_SIZE, %rdi + VPCMP $0, (%rdi), %YMMZERO, %k0 + kmovd %k0, %eax + /* Remove the leading bytes. */ # ifdef USE_AS_WCSLEN - shrq $2, %rax + /* NB: Divide shift count by 4 since each bit in K0 represent 4 + bytes. */ + movl %edx, %ecx + shrl $2, %ecx + andl $(CHAR_PER_VEC - 1), %ecx # endif - ret - - .p2align 4 -L(first_vec_x1): + /* SHIFT_REG is ecx for USE_AS_WCSLEN and edx otherwise. */ + sarxl %SHIFT_REG, %eax, %eax + testl %eax, %eax +# ifndef USE_AS_STRNLEN + jz L(cross_page_continue) tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - addq $VEC_SIZE, %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif ret - - .p2align 4 -L(first_vec_x2): - tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - addq $(VEC_SIZE * 2), %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif +# else + jnz L(cross_page_less_vec) +# ifndef USE_AS_WCSLEN + movl %edx, %ecx + andl $(CHAR_PER_VEC - 1), %ecx +# endif + movl $CHAR_PER_VEC, %eax + subl %ecx, %eax + /* Check the end of data. */ + cmpq %rax, %rsi + ja L(cross_page_continue) + movl %esi, %eax ret - - .p2align 4 -L(4x_vec_end): - VPCMP $0, %YMM1, %YMMZERO, %k0 - kmovd %k0, %eax - testl %eax, %eax - jnz L(first_vec_x0) - VPCMP $0, %YMM2, %YMMZERO, %k1 - kmovd %k1, %eax - testl %eax, %eax - jnz L(first_vec_x1) - VPCMP $0, %YMM3, %YMMZERO, %k2 - kmovd %k2, %eax - testl %eax, %eax - jnz L(first_vec_x2) - VPCMP $0, %YMM4, %YMMZERO, %k3 - kmovd %k3, %eax -L(first_vec_x3): +L(cross_page_less_vec): tzcntl %eax, %eax -# ifdef USE_AS_WCSLEN - /* NB: Multiply wchar_t count by 4 to get the number of bytes. */ - sall $2, %eax -# endif - addq $(VEC_SIZE * 3), %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif + /* Select min of length and position of first null. */ + cmpq %rax, %rsi + cmovb %esi, %eax ret +# endif END (STRLEN) #endif From patchwork Mon Apr 19 16:28:59 2021 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Noah Goldstein X-Patchwork-Id: 43042 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 A71D33956C2E; Mon, 19 Apr 2021 16:29:13 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org A71D33956C2E DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1618849753; bh=/3CyeNFh15RVuwfWeoolxaeXqbCJaNy+vM78WWHiWRM=; 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=Lmkt0q92vj6u3M5zOI2H4Jc7Na4aAaTwmBj6ssodKN/AcBE7hieVr2dHUqSyw+Bao oOrAAbgj4mgc/FVDDtzedcNKyTAyOLVwTDgDAkOxR/KmPTLjlChcZRDUo93/y6z+bl LdVQ0a3AbEn8CxggCmMRdJPzspgydNXoIZzrE8v8= X-Original-To: libc-alpha@sourceware.org Delivered-To: libc-alpha@sourceware.org Received: from mail-qv1-xf31.google.com (mail-qv1-xf31.google.com [IPv6:2607:f8b0:4864:20::f31]) by sourceware.org (Postfix) with ESMTPS id 700EA39518BE for ; Mon, 19 Apr 2021 16:29:09 +0000 (GMT) DMARC-Filter: OpenDMARC Filter v1.3.2 sourceware.org 700EA39518BE Received: by mail-qv1-xf31.google.com with SMTP id j3so17163672qvs.1 for ; Mon, 19 Apr 2021 09:29:09 -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=/3CyeNFh15RVuwfWeoolxaeXqbCJaNy+vM78WWHiWRM=; b=CvcE6FGo5XLlqWqKyVM1FH4XlS/9o5rHGmP20Uw9QStQnVz97EBrO3njFbIhUadSrN YCv12lPBggJNpGfwu1kPU6CJhuFjLWX68xDavZIfvsA7eACLac/fvGWv7e1yxi7zBY3M jqoT0AXQIBsYzTuPvG8L+DR+Phf0ZiSJk620VZjmJsyG1nWJ7tBThnBitTl/9oYDWH4C oZrxmlUrUNhwnbmBalbF+qSsY1cnqTfv2hZyu7K9Cu8BN+8wL5LI8kqPXHhajrOcrNzE 87YiT2jXo2h64b8SQIQlv/HyX8Xpo8QAempR+jtqeAt9i8y81KUZhCGjiqk3JRrL5HT6 SQ6A== X-Gm-Message-State: AOAM5309GgcHrnG3jx8qjGOCZ3FwqbqRhuYValGb/oq1UZGiwdvg2OvM 7vSgmZPzCTi/khRdKWXJgovErbb1r9I= X-Google-Smtp-Source: ABdhPJxsWj/WJUIhldWZwKGhTOXwkL6LouhiJ4YN9/W3x4o7S9Cj4BaZBYoganJzoL8x25QGap4rLw== X-Received: by 2002:a0c:ee81:: with SMTP id u1mr21813626qvr.14.1618849748506; Mon, 19 Apr 2021 09:29:08 -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 c23sm9573332qtm.46.2021.04.19.09.29.07 (version=TLS1_3 cipher=TLS_AES_256_GCM_SHA384 bits=256/256); Mon, 19 Apr 2021 09:29:08 -0700 (PDT) To: libc-alpha@sourceware.org Subject: [PATCH v3 2/2] x86: Optimize strlen-avx2.S Date: Mon, 19 Apr 2021 12:28:59 -0400 Message-Id: <20210419162859.2285431-2-goldstein.w.n@gmail.com> X-Mailer: git-send-email 2.29.2 In-Reply-To: <20210419162859.2285431-1-goldstein.w.n@gmail.com> References: <20210419162859.2285431-1-goldstein.w.n@gmail.com> MIME-Version: 1.0 X-Spam-Status: No, score=-12.5 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 strlen-avx2.S. The optimizations are mostly small things but they add up to roughly 10-30% performance improvement for strlen. The results for strnlen are bit more ambiguous. test-strlen, test-strnlen, test-wcslen, and test-wcsnlen are all passing. Signed-off-by: Noah Goldstein --- sysdeps/x86_64/multiarch/ifunc-avx2.h | 4 +- sysdeps/x86_64/multiarch/ifunc-impl-list.c | 14 +- sysdeps/x86_64/multiarch/strlen-avx2.S | 535 +++++++++++++-------- 3 files changed, 337 insertions(+), 216 deletions(-) diff --git a/sysdeps/x86_64/multiarch/ifunc-avx2.h b/sysdeps/x86_64/multiarch/ifunc-avx2.h index e3ec62ca5e..6de72f7272 100644 --- a/sysdeps/x86_64/multiarch/ifunc-avx2.h +++ b/sysdeps/x86_64/multiarch/ifunc-avx2.h @@ -30,11 +30,11 @@ IFUNC_SELECTOR (void) const struct cpu_features* cpu_features = __get_cpu_features (); if (CPU_FEATURE_USABLE_P (cpu_features, AVX2) + && CPU_FEATURE_USABLE_P (cpu_features, BMI2) && CPU_FEATURES_ARCH_P (cpu_features, AVX_Fast_Unaligned_Load)) { if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL) - && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW) - && CPU_FEATURE_USABLE_P (cpu_features, BMI2)) + && CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)) return OPTIMIZE (evex); if (CPU_FEATURE_USABLE_P (cpu_features, RTM)) diff --git a/sysdeps/x86_64/multiarch/ifunc-impl-list.c b/sysdeps/x86_64/multiarch/ifunc-impl-list.c index 770986e41a..4c992e0cc5 100644 --- a/sysdeps/x86_64/multiarch/ifunc-impl-list.c +++ b/sysdeps/x86_64/multiarch/ifunc-impl-list.c @@ -285,30 +285,36 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array, /* Support sysdeps/x86_64/multiarch/strlen.c. */ IFUNC_IMPL (i, name, strlen, IFUNC_IMPL_ADD (array, i, strlen, - CPU_FEATURE_USABLE (AVX2), + (CPU_FEATURE_USABLE (AVX2) + && CPU_FEATURE_USABLE (BMI2)), __strlen_avx2) IFUNC_IMPL_ADD (array, i, strlen, (CPU_FEATURE_USABLE (AVX2) + && CPU_FEATURE_USABLE (BMI2) && CPU_FEATURE_USABLE (RTM)), __strlen_avx2_rtm) IFUNC_IMPL_ADD (array, i, strlen, (CPU_FEATURE_USABLE (AVX512VL) - && CPU_FEATURE_USABLE (AVX512BW)), + && CPU_FEATURE_USABLE (AVX512BW) + && CPU_FEATURE_USABLE (BMI2)), __strlen_evex) IFUNC_IMPL_ADD (array, i, strlen, 1, __strlen_sse2)) /* Support sysdeps/x86_64/multiarch/strnlen.c. */ IFUNC_IMPL (i, name, strnlen, IFUNC_IMPL_ADD (array, i, strnlen, - CPU_FEATURE_USABLE (AVX2), + (CPU_FEATURE_USABLE (AVX2) + && CPU_FEATURE_USABLE (BMI2)), __strnlen_avx2) IFUNC_IMPL_ADD (array, i, strnlen, (CPU_FEATURE_USABLE (AVX2) + && CPU_FEATURE_USABLE (BMI2) && CPU_FEATURE_USABLE (RTM)), __strnlen_avx2_rtm) IFUNC_IMPL_ADD (array, i, strnlen, (CPU_FEATURE_USABLE (AVX512VL) - && CPU_FEATURE_USABLE (AVX512BW)), + && CPU_FEATURE_USABLE (AVX512BW) + && CPU_FEATURE_USABLE (BMI2)), __strnlen_evex) IFUNC_IMPL_ADD (array, i, strnlen, 1, __strnlen_sse2)) diff --git a/sysdeps/x86_64/multiarch/strlen-avx2.S b/sysdeps/x86_64/multiarch/strlen-avx2.S index 1caae9e6bc..eb92f587b4 100644 --- a/sysdeps/x86_64/multiarch/strlen-avx2.S +++ b/sysdeps/x86_64/multiarch/strlen-avx2.S @@ -27,9 +27,11 @@ # ifdef USE_AS_WCSLEN # define VPCMPEQ vpcmpeqd # define VPMINU vpminud +# define CHAR_SIZE 4 # else # define VPCMPEQ vpcmpeqb # define VPMINU vpminub +# define CHAR_SIZE 1 # endif # ifndef VZEROUPPER @@ -41,349 +43,462 @@ # endif # define VEC_SIZE 32 +# define PAGE_SIZE 4096 +# define LOG_PAGE_SIZE 12 .section SECTION(.text),"ax",@progbits ENTRY (STRLEN) # ifdef USE_AS_STRNLEN - /* Check for zero length. */ + /* Check zero length. */ test %RSI_LP, %RSI_LP jz L(zero) + /* Store max len in R8_LP before adjusting if using WCSLEN. */ + mov %RSI_LP, %R8_LP # ifdef USE_AS_WCSLEN shl $2, %RSI_LP # elif defined __ILP32__ /* Clear the upper 32 bits. */ movl %esi, %esi # endif - mov %RSI_LP, %R8_LP # endif - movl %edi, %ecx + movl %edi, %eax movq %rdi, %rdx vpxor %xmm0, %xmm0, %xmm0 - + /* Shift left eax to clear all bits not relevant to page cross + check. This saves 2 bytes of code as opposed to using andl with + PAGE_SIZE - 1. Then compare with PAGE_SIZE - VEC_SIZE shifted + left by the same amount (an imm32 either way). */ + sall $(32 - LOG_PAGE_SIZE), %eax /* 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) + cmpl $((PAGE_SIZE - VEC_SIZE) << (32 - LOG_PAGE_SIZE)), %eax + ja L(cross_page_boundary) /* Check the first VEC_SIZE bytes. */ - VPCMPEQ (%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - + VPCMPEQ (%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax # ifdef USE_AS_STRNLEN - jnz L(first_vec_x0_check) - /* Adjust length and check the end of data. */ - subq $VEC_SIZE, %rsi - jbe L(max) -# else - jnz L(first_vec_x0) + /* If length < VEC_SIZE handle special. */ + cmpq $VEC_SIZE, %rsi + 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 + /* If empty continue to aligned_more. Otherwise return bit + position of first match. */ + testl %eax, %eax + jz L(aligned_more) + tzcntl %eax, %eax +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif + VZEROUPPER_RETURN # ifdef USE_AS_STRNLEN - /* Adjust length. */ - addq %rcx, %rsi +L(zero): + xorl %eax, %eax + ret - subq $(VEC_SIZE * 4), %rsi - jbe L(last_4x_vec_or_less) + .p2align 4 +L(first_vec_x0): + /* Set bit for max len so that tzcnt will return min of max len + and position of first match. */ + btsq %rsi, %rax + tzcntl %eax, %eax +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif + VZEROUPPER_RETURN # 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 - vpmovmskb %ymm1, %eax - /* Remove the leading bytes. */ - sarl %cl, %eax - testl %eax, %eax - jz L(aligned_more) +L(first_vec_x1): tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) + /* Use ecx which was computed earlier to compute correct value. + */ + subl $(VEC_SIZE * 4 + 1), %ecx + addl %ecx, %eax +# else + subl %edx, %edi + incl %edi + addl %edi, %eax # endif - addq %rdi, %rax - addq %rcx, %rax - subq %rdx, %rax # ifdef USE_AS_WCSLEN - shrq $2, %rax + shrl $2, %eax # endif -L(return_vzeroupper): - ZERO_UPPER_VEC_REGISTERS_RETURN + VZEROUPPER_RETURN .p2align 4 -L(aligned_more): +L(first_vec_x2): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - /* "rcx" is less than VEC_SIZE. 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 - - /* Check the end of data. */ - subq %rcx, %rsi - jbe L(max) + /* Use ecx which was computed earlier to compute correct value. + */ + subl $(VEC_SIZE * 3 + 1), %ecx + addl %ecx, %eax +# else + subl %edx, %edi + addl $(VEC_SIZE + 1), %edi + addl %edi, %eax # endif +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif + VZEROUPPER_RETURN - addq $VEC_SIZE, %rdi + .p2align 4 +L(first_vec_x3): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ +# ifdef USE_AS_STRNLEN + /* Use ecx which was computed earlier to compute correct value. + */ + subl $(VEC_SIZE * 2 + 1), %ecx + addl %ecx, %eax +# else + subl %edx, %edi + addl $(VEC_SIZE * 2 + 1), %edi + addl %edi, %eax +# endif +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif + VZEROUPPER_RETURN + .p2align 4 +L(first_vec_x4): + tzcntl %eax, %eax + /* Safe to use 32 bit instructions as these are only called for + size = [1, 159]. */ # ifdef USE_AS_STRNLEN - subq $(VEC_SIZE * 4), %rsi - jbe L(last_4x_vec_or_less) + /* Use ecx which was computed earlier to compute correct value. + */ + subl $(VEC_SIZE + 1), %ecx + addl %ecx, %eax +# else + subl %edx, %edi + addl $(VEC_SIZE * 3 + 1), %edi + addl %edi, %eax # endif +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif + VZEROUPPER_RETURN -L(more_4x_vec): + .p2align 5 +L(aligned_more): + /* Align data to VEC_SIZE - 1. This is the same number of + instructions as using andq with -VEC_SIZE but saves 4 bytes of + code on the x4 check. */ + orq $(VEC_SIZE - 1), %rdi +L(cross_page_continue): /* 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) - - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax +# ifdef USE_AS_STRNLEN + /* + 1 because rdi is aligned to VEC_SIZE - 1. + CHAR_SIZE because + it simplies the logic in last_4x_vec_or_less. */ + leaq (VEC_SIZE * 4 + CHAR_SIZE + 1)(%rdi), %rcx + subq %rdx, %rcx +# endif + /* Load first VEC regardless. */ + VPCMPEQ 1(%rdi), %ymm0, %ymm1 +# ifdef USE_AS_STRNLEN + /* Adjust length. If near end handle specially. */ + subq %rcx, %rsi + jb L(last_4x_vec_or_less) +# endif + vpmovmskb %ymm1, %eax testl %eax, %eax jnz L(first_vec_x1) - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax + 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 - vpmovmskb %ymm1, %eax + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax testl %eax, %eax jnz L(first_vec_x3) - addq $(VEC_SIZE * 4), %rdi - -# ifdef USE_AS_STRNLEN - subq $(VEC_SIZE * 4), %rsi - 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 + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + testl %eax, %eax + jnz L(first_vec_x4) + /* Align data to VEC_SIZE * 4 - 1. */ # ifdef USE_AS_STRNLEN - /* Adjust length. */ + /* Before adjusting length check if at last VEC_SIZE * 4. */ + cmpq $(VEC_SIZE * 4 - 1), %rsi + jbe L(last_4x_vec_or_less_load) + incq %rdi + movl %edi, %ecx + orq $(VEC_SIZE * 4 - 1), %rdi + andl $(VEC_SIZE * 4 - 1), %ecx + /* Readjust length. */ addq %rcx, %rsi +# else + 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. */ - vmovdqa (%rdi), %ymm1 - vmovdqa VEC_SIZE(%rdi), %ymm2 - vmovdqa (VEC_SIZE * 2)(%rdi), %ymm3 - vmovdqa (VEC_SIZE * 3)(%rdi), %ymm4 - VPMINU %ymm1, %ymm2, %ymm5 - VPMINU %ymm3, %ymm4, %ymm6 - VPMINU %ymm5, %ymm6, %ymm5 - - VPCMPEQ %ymm5, %ymm0, %ymm5 - vpmovmskb %ymm5, %eax - testl %eax, %eax - jnz L(4x_vec_end) - - addq $(VEC_SIZE * 4), %rdi - -# ifndef USE_AS_STRNLEN - jmp L(loop_4x_vec) -# else +# ifdef USE_AS_STRNLEN + /* Break if at end of length. */ subq $(VEC_SIZE * 4), %rsi - ja L(loop_4x_vec) - -L(last_4x_vec_or_less): - /* Less than 4 * VEC and aligned to VEC_SIZE. */ - addl $(VEC_SIZE * 2), %esi - jle L(last_2x_vec) + jb L(last_4x_vec_or_less_cmpeq) +# endif + /* Save some code size by microfusing VPMINU with the load. Since + the matches in ymm2/ymm4 can only be returned if there where no + matches in ymm1/ymm3 respectively there is no issue with overlap. + */ + vmovdqa 1(%rdi), %ymm1 + VPMINU (VEC_SIZE + 1)(%rdi), %ymm1, %ymm2 + vmovdqa (VEC_SIZE * 2 + 1)(%rdi), %ymm3 + VPMINU (VEC_SIZE * 3 + 1)(%rdi), %ymm3, %ymm4 + + VPMINU %ymm2, %ymm4, %ymm5 + VPCMPEQ %ymm5, %ymm0, %ymm5 + vpmovmskb %ymm5, %ecx - VPCMPEQ (%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x0) + subq $-(VEC_SIZE * 4), %rdi + testl %ecx, %ecx + jz L(loop_4x_vec) - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x1) - VPCMPEQ (VEC_SIZE * 2)(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax + VPCMPEQ %ymm1, %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + subq %rdx, %rdi testl %eax, %eax + jnz L(last_vec_return_x0) - jnz L(first_vec_x2_check) - subl $VEC_SIZE, %esi - jle L(max) - - VPCMPEQ (VEC_SIZE * 3)(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax + VPCMPEQ %ymm2, %ymm0, %ymm2 + vpmovmskb %ymm2, %eax testl %eax, %eax - - jnz L(first_vec_x3_check) - movq %r8, %rax -# ifdef USE_AS_WCSLEN + jnz L(last_vec_return_x1) + + /* Combine last 2 VEC. */ + VPCMPEQ %ymm3, %ymm0, %ymm3 + vpmovmskb %ymm3, %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 - 1), %rdi + addq %rdi, %rax +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif VZEROUPPER_RETURN + +# ifdef USE_AS_STRNLEN .p2align 4 -L(last_2x_vec): - addl $(VEC_SIZE * 2), %esi - VPCMPEQ (%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax +L(last_4x_vec_or_less_load): + /* Depending on entry adjust rdi / prepare first VEC in ymm1. */ + subq $-(VEC_SIZE * 4), %rdi +L(last_4x_vec_or_less_cmpeq): + VPCMPEQ 1(%rdi), %ymm0, %ymm1 +L(last_4x_vec_or_less): - jnz L(first_vec_x0_check) - subl $VEC_SIZE, %esi - jle L(max) + vpmovmskb %ymm1, %eax + /* If remaining length > VEC_SIZE * 2. This works if esi is off by + VEC_SIZE * 4. */ + testl $(VEC_SIZE * 2), %esi + jnz L(last_4x_vec) - VPCMPEQ VEC_SIZE(%rdi), %ymm0, %ymm1 - vpmovmskb %ymm1, %eax + /* length may have been negative or positive by an offset of + VEC_SIZE * 4 depending on where this was called from. This fixes + that. */ + andl $(VEC_SIZE * 4 - 1), %esi testl %eax, %eax - jnz L(first_vec_x1_check) - movq %r8, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax -# endif - VZEROUPPER_RETURN + jnz L(last_vec_x1_check) - .p2align 4 -L(first_vec_x0_check): + subl $VEC_SIZE, %esi + jb L(max) + + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax tzcntl %eax, %eax /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) + cmpl %eax, %esi + jb L(max) + subq %rdx, %rdi + addl $(VEC_SIZE + 1), %eax addq %rdi, %rax - subq %rdx, %rax # ifdef USE_AS_WCSLEN shrq $2, %rax # endif VZEROUPPER_RETURN +# endif .p2align 4 -L(first_vec_x1_check): +L(last_vec_return_x0): tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $VEC_SIZE, %rax + subq $(VEC_SIZE * 4 - 1), %rdi addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif VZEROUPPER_RETURN .p2align 4 -L(first_vec_x2_check): +L(last_vec_return_x1): tzcntl %eax, %eax - /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $(VEC_SIZE * 2), %rax + subq $(VEC_SIZE * 3 - 1), %rdi addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif VZEROUPPER_RETURN +# ifdef USE_AS_STRNLEN .p2align 4 -L(first_vec_x3_check): +L(last_vec_x1_check): + tzcntl %eax, %eax /* Check the end of data. */ - cmpq %rax, %rsi - jbe L(max) - addq $(VEC_SIZE * 3), %rax + cmpl %eax, %esi + jb L(max) + subq %rdx, %rdi + incl %eax addq %rdi, %rax - subq %rdx, %rax # ifdef USE_AS_WCSLEN shrq $2, %rax # endif VZEROUPPER_RETURN - .p2align 4 L(max): movq %r8, %rax + VZEROUPPER_RETURN + + .p2align 4 +L(last_4x_vec): + /* Test first 2x VEC normally. */ + testl %eax, %eax + jnz L(last_vec_x1) + + VPCMPEQ (VEC_SIZE + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + testl %eax, %eax + jnz L(last_vec_x2) + + /* Normalize length. */ + andl $(VEC_SIZE * 4 - 1), %esi + VPCMPEQ (VEC_SIZE * 2 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + testl %eax, %eax + jnz L(last_vec_x3) + + subl $(VEC_SIZE * 3), %esi + jb L(max) + + VPCMPEQ (VEC_SIZE * 3 + 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + tzcntl %eax, %eax + /* Check the end of data. */ + cmpl %eax, %esi + jb L(max) + subq %rdx, %rdi + addl $(VEC_SIZE * 3 + 1), %eax + addq %rdi, %rax # ifdef USE_AS_WCSLEN shrq $2, %rax # endif VZEROUPPER_RETURN - .p2align 4 -L(zero): - xorl %eax, %eax - ret -# endif .p2align 4 -L(first_vec_x0): +L(last_vec_x1): + /* essentially duplicates of first_vec_x1 but use 64 bit + instructions. */ tzcntl %eax, %eax + subq %rdx, %rdi + incl %eax addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif VZEROUPPER_RETURN .p2align 4 -L(first_vec_x1): +L(last_vec_x2): + /* essentially duplicates of first_vec_x1 but use 64 bit + instructions. */ tzcntl %eax, %eax - addq $VEC_SIZE, %rax + subq %rdx, %rdi + addl $(VEC_SIZE + 1), %eax addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif VZEROUPPER_RETURN .p2align 4 -L(first_vec_x2): +L(last_vec_x3): tzcntl %eax, %eax - addq $(VEC_SIZE * 2), %rax + subl $(VEC_SIZE * 2), %esi + /* Check the end of data. */ + cmpl %eax, %esi + jb L(max_end) + subq %rdx, %rdi + addl $(VEC_SIZE * 2 + 1), %eax addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN +# ifdef USE_AS_WCSLEN shrq $2, %rax -# endif +# endif + VZEROUPPER_RETURN +L(max_end): + movq %r8, %rax VZEROUPPER_RETURN +# endif + /* Cold case for crossing page with first load. */ .p2align 4 -L(4x_vec_end): - VPCMPEQ %ymm1, %ymm0, %ymm1 - vpmovmskb %ymm1, %eax - testl %eax, %eax - jnz L(first_vec_x0) - VPCMPEQ %ymm2, %ymm0, %ymm2 - vpmovmskb %ymm2, %eax +L(cross_page_boundary): + /* Align data to VEC_SIZE - 1. */ + orq $(VEC_SIZE - 1), %rdi + VPCMPEQ -(VEC_SIZE - 1)(%rdi), %ymm0, %ymm1 + vpmovmskb %ymm1, %eax + /* Remove the leading bytes. sarxl only uses bits [5:0] of COUNT + so no need to manually mod rdx. */ + sarxl %edx, %eax, %eax +# ifdef USE_AS_STRNLEN testl %eax, %eax - jnz L(first_vec_x1) - VPCMPEQ %ymm3, %ymm0, %ymm3 - vpmovmskb %ymm3, %eax + jnz L(cross_page_less_vec) + leaq 1(%rdi), %rcx + subq %rdx, %rcx + /* Check length. */ + cmpq %rsi, %rcx + jb L(cross_page_continue) + movq %r8, %rax +# else testl %eax, %eax - jnz L(first_vec_x2) - VPCMPEQ %ymm4, %ymm0, %ymm4 - vpmovmskb %ymm4, %eax -L(first_vec_x3): + jz L(cross_page_continue) tzcntl %eax, %eax - addq $(VEC_SIZE * 3), %rax - addq %rdi, %rax - subq %rdx, %rax -# ifdef USE_AS_WCSLEN - shrq $2, %rax +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif # endif +L(return_vzeroupper): + ZERO_UPPER_VEC_REGISTERS_RETURN + +# ifdef USE_AS_STRNLEN + .p2align 4 +L(cross_page_less_vec): + tzcntl %eax, %eax + cmpq %rax, %rsi + cmovb %esi, %eax +# ifdef USE_AS_WCSLEN + shrl $2, %eax +# endif VZEROUPPER_RETURN +# endif END (STRLEN) #endif