From patchwork Tue May 21 13:12:21 2019 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Wilco Dijkstra X-Patchwork-Id: 32781 Received: (qmail 17140 invoked by alias); 21 May 2019 13:12:28 -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 17132 invoked by uid 89); 21 May 2019 13:12:27 -0000 Authentication-Results: sourceware.org; auth=none X-Spam-SWARE-Status: No, score=-17.2 required=5.0 tests=AWL, BAYES_00, GIT_PATCH_0, GIT_PATCH_1, GIT_PATCH_2, GIT_PATCH_3, RCVD_IN_DNSWL_NONE, SPF_HELO_PASS, SPF_PASS autolearn=ham version=3.3.1 spammy= X-HELO: EUR04-HE1-obe.outbound.protection.outlook.com DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=armh.onmicrosoft.com; s=selector2-armh-onmicrosoft-com; h=From:Date:Subject:Message-ID:Content-Type:MIME-Version:X-MS-Exchange-SenderADCheck; bh=pM0WKuAmyzYliDUvjh7taPWiHmOkth4TuSmXiTfXcn8=; b=8yOHuyR9q282ZSRHH8cEyYozOlYZs0JE4rTjhErhQRahpF+JHyFx8RD6/ja44IAx3XxEyj/e/ub/zw7ByOtNW044Q4C4k20EgUejJR+hUd5BUoNzypsXZJSxgS54qjrdcUTXEPTNvosE2FxaxLuBkuNQ9WC1Nzim/AJa67kS5Uk= From: Wilco Dijkstra To: Adhemerval Zanella , "libc-alpha@sourceware.org" CC: nd Subject: Re: [PATCH v2] Benchmark strstr hard needles Date: Tue, 21 May 2019 13:12:21 +0000 Message-ID: References: , <3cacddd7-308b-7579-e6f8-50f6c49778cc@linaro.org> In-Reply-To: <3cacddd7-308b-7579-e6f8-50f6c49778cc@linaro.org> authentication-results: spf=none (sender IP is ) smtp.mailfrom=Wilco.Dijkstra@arm.com; x-ms-oob-tlc-oobclassifiers: OLM:2512; received-spf: None (protection.outlook.com: arm.com does not designate permitted sender hosts) x-ms-exchange-senderadcheck: 1 MIME-Version: 1.0 X-MS-Exchange-CrossTenant-mailboxtype: HOSTED Hi Adhemerval, > +char * volatile null = NULL; > I am not seeing why exactly it requires a global non static volatile > variable to call do_one_test. You're right, this is not necessary in the current version (IIRC it was to avoid a warning), so I have removed it. > Maybe add some context from where these tests came from, from the > libc-alpha discussion? Sure, I've improved the comments and description to make it clearer what the hard needle test is doing and why. New version below. Wilco [PATCH v3] Benchmark strstr hard needles v3: Improve comments/description. Benchmark needles which exhibit worst-case performance. This shows that basic_strstr is quadratic and thus unsuitable for large needles. On the other hand the Two-way and new strstr implementations are linear with increasing needle sizes. The slowest cases of the two implementations are within a factor of 2 on several different microarchitectures. Two-way is slowest on inputs which cause a branch mispredict on almost every character. The new strstr is slowest on inputs which almost match and result in many calls to memcmp. Thanks to Szabolcs for providing various hard needles. ChangeLog: 2019-05-21 Wilco Dijkstra * benchtests/bench-strstr.c (test_hard_needle): New function. diff --git a/benchtests/bench-strstr.c b/benchtests/bench-strstr.c index 31309b24029a96de7381e1050fd89e5d26642e5f..5ec6edcf1869dd198a3326cb85f2a4463ae3417c 100644 --- a/benchtests/bench-strstr.c +++ b/benchtests/bench-strstr.c @@ -203,6 +203,81 @@ do_test (size_t align1, size_t align2, size_t len1, size_t len2, putchar ('\n'); } +/* Test needles which exhibit worst-case performance. This shows that + basic_strstr is quadratic and thus unsuitable for large needles. + On the other hand Two-way and skip table implementations are linear with + increasing needle sizes. The slowest cases of the two implementations are + within a factor of 2 on several different microarchitectures. */ + +static void +test_hard_needle (size_t ne_len, size_t hs_len) +{ + char *ne = (char *) buf1; + char *hs = (char *) buf2; + + /* Hard needle for strstr algorithm using skip table. This results in many + memcmp calls comparing most of the needle. */ + { + memset (ne, 'a', ne_len); + ne[ne_len] = '\0'; + ne[ne_len - 14] = 'b'; + + memset (hs, 'a', hs_len); + for (size_t i = ne_len; i <= hs_len; i += ne_len) + { + hs[i-5] = 'b'; + hs[i-62] = 'b'; + } + + printf ("Length %4zd/%3zd, complex needle 1:", hs_len, ne_len); + + FOR_EACH_IMPL (impl, 0) + do_one_test (impl, hs, ne, NULL); + putchar ('\n'); + } + + /* 2nd hard needle for strstr algorithm using skip table. This results in + many memcmp calls comparing most of the needle. */ + { + memset (ne, 'a', ne_len); + ne[ne_len] = '\0'; + ne[ne_len - 6] = 'b'; + + memset (hs, 'a', hs_len); + for (size_t i = ne_len; i <= hs_len; i += ne_len) + { + hs[i-5] = 'b'; + hs[i-6] = 'b'; + } + + printf ("Length %4zd/%3zd, complex needle 2:", hs_len, ne_len); + + FOR_EACH_IMPL (impl, 0) + do_one_test (impl, hs, ne, NULL); + putchar ('\n'); + } + + /* Hard needle for Two-way algorithm - the random input causes a large number + of branch mispredictions which significantly reduces performance on modern + micro architectures. */ + { + for (int i = 0; i < hs_len; i++) + hs[i] = (rand () & 255) > 155 ? 'a' : 'b'; + hs[hs_len] = 0; + + memset (ne, 'a', ne_len); + ne[ne_len-2] = 'b'; + ne[0] = 'b'; + ne[ne_len] = 0; + + printf ("Length %4zd/%3zd, complex needle 3:", hs_len, ne_len); + + FOR_EACH_IMPL (impl, 0) + do_one_test (impl, hs, ne, NULL); + putchar ('\n'); + } +} + static int test_main (void) { @@ -227,6 +302,10 @@ test_main (void) do_test (14, 5, hlen, klen, 1); } + test_hard_needle (64, 65536); + test_hard_needle (256, 65536); + test_hard_needle (1024, 65536); + return ret; }