From patchwork Fri Feb 2 04:50:56 2018 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Siddhesh Poyarekar X-Patchwork-Id: 25762 Received: (qmail 52053 invoked by alias); 2 Feb 2018 04:51:29 -0000 Mailing-List: contact libc-alpha-help@sourceware.org; run by ezmlm Precedence: bulk List-Id: List-Unsubscribe: List-Subscribe: List-Archive: List-Post: List-Help: , Sender: libc-alpha-owner@sourceware.org Delivered-To: mailing list libc-alpha@sourceware.org Received: (qmail 52044 invoked by uid 89); 2 Feb 2018 04:51:29 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-26.1 required=5.0 tests=BAYES_00, GIT_PATCH_0, GIT_PATCH_1, GIT_PATCH_2, GIT_PATCH_3, RCVD_IN_DNSWL_NONE, SPF_NEUTRAL autolearn=ham version=3.3.2 spammy=UD:gt X-HELO: homiemail-a52.g.dreamhost.com From: Siddhesh Poyarekar To: libc-alpha@sourceware.org Cc: szabolcs.nagy@arm.com Subject: [PATCH] aarch64: Optimized memcmp for medium to large sizes Date: Fri, 2 Feb 2018 10:20:56 +0530 Message-Id: <20180202045056.3121-1-siddhesh@sourceware.org> This improved memcmp provides a fast path for compares up to 16 bytes and then compares 16 bytes at a time, thus optimizing loads from both sources. The glibc memcmp microbenchmark retains performance (with an error of ~1ns) for smaller compare sizes and reduces up to 31% of execution time for compares up to 4K on the APM Mustang. On Qualcomm Falkor this improves to almost 48%, i.e. it is almost 2x improvement for sizes of 2K and above. * sysdeps/aarch64/memcmp.S: Widen comparison to 16 bytes at a time. --- sysdeps/aarch64/memcmp.S | 66 ++++++++++++++++++++++++++++++++++++------------ 1 file changed, 50 insertions(+), 16 deletions(-) diff --git a/sysdeps/aarch64/memcmp.S b/sysdeps/aarch64/memcmp.S index ecd12061b2..0804e75d99 100644 --- a/sysdeps/aarch64/memcmp.S +++ b/sysdeps/aarch64/memcmp.S @@ -34,9 +34,12 @@ /* Internal variables. */ #define data1 x3 #define data1w w3 -#define data2 x4 -#define data2w w4 -#define tmp1 x5 +#define data1h x4 +#define data2 x5 +#define data2w w5 +#define data2h x6 +#define tmp1 x7 +#define tmp2 x8 ENTRY_ALIGN (memcmp, 6) DELOUSE (0) @@ -46,39 +49,70 @@ ENTRY_ALIGN (memcmp, 6) subs limit, limit, 8 b.lo L(less8) - /* Limit >= 8, so check first 8 bytes using unaligned loads. */ ldr data1, [src1], 8 ldr data2, [src2], 8 - and tmp1, src1, 7 - add limit, limit, tmp1 + cmp data1, data2 + b.ne L(return) + + subs limit, limit, 8 + b.gt L(more16) + + ldr data1, [src1, limit] + ldr data2, [src2, limit] + b L(return) + +L(more16): + ldr data1, [src1], 8 + ldr data2, [src2], 8 cmp data1, data2 bne L(return) + /* Jump directly to copying the last 16 bytes for 32 byte (or less) + strings. */ + subs limit, limit, 16 + b.ls L(last_bytes) + + /* We overlap loads between 0-32 bytes at either side of SRC1 when we + try to align, so limit it only to strings larger than 128 bytes. */ + cmp limit, 96 + b.ls L(loop8) + /* Align src1 and adjust src2 with bytes not yet done. */ + and tmp1, src1, 15 + add limit, limit, tmp1 sub src1, src1, tmp1 sub src2, src2, tmp1 - subs limit, limit, 8 - b.ls L(last_bytes) - /* Loop performing 8 bytes per iteration using aligned src1. Limit is pre-decremented by 8 and must be larger than zero. Exit if <= 8 bytes left to do or if the data is not equal. */ .p2align 4 L(loop8): - ldr data1, [src1], 8 - ldr data2, [src2], 8 - subs limit, limit, 8 - ccmp data1, data2, 0, hi /* NZCV = 0b0000. */ + ldp data1, data1h, [src1], 16 + ldp data2, data2h, [src2], 16 + subs limit, limit, 16 + ccmp data1, data2, 0, hi + ccmp data1h, data2h, 0, eq b.eq L(loop8) + cmp data1, data2 + bne L(return) + mov data1, data1h + mov data2, data2h cmp data1, data2 bne L(return) - /* Compare last 1-8 bytes using unaligned access. */ + /* Compare last 1-16 bytes using unaligned access. */ L(last_bytes): - ldr data1, [src1, limit] - ldr data2, [src2, limit] + add src1, src1, limit + add src2, src2, limit + ldp data1, data1h, [src1] + ldp data2, data2h, [src2] + cmp data1, data2 + bne L(return) + mov data1, data1h + mov data2, data2h + cmp data1, data2 /* Compare data bytes and set return value to 0, -1 or 1. */ L(return):