From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: by sourceware.org (Postfix, from userid 1895) id 91D213888C7F; Wed, 10 Apr 2024 17:18:56 +0000 (GMT) DKIM-Filter: OpenDKIM Filter v2.11.0 sourceware.org 91D213888C7F DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=sourceware.org; s=default; t=1712769536; bh=bTWPH3hl1gznJU4P/OX18aV2SayfEM1rMHCw34vm4hQ=; h=From:To:Subject:Date:From; b=fEK5ss5FdQZzqCT8SDUB9U+Lr0XqQ1/pbtF2SrCYwzt3s1bCHPQoUG/Z44f5KwvXa bmjfaUviftyBB/7O5ICY9YZGEKtLX8wNm2OP3PqgFxfoDDZXyvnxF9/jxsq2bWC499 q79nqp479lutme8EA6KFFlcyErxPfkH+Ub4IcNrw= Content-Type: text/plain; charset="us-ascii" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit From: Wilco Dijkstra To: glibc-cvs@sourceware.org Subject: [glibc/release/2.32/master] AArch64: Optimize memrchr X-Act-Checkin: glibc X-Git-Author: Wilco Dijkstra X-Git-Refname: refs/heads/release/2.32/master X-Git-Oldrev: 91680682e59a855e5315586bae0b5c228a785aac X-Git-Newrev: f55ba2fedc90438dcba7d3fdc4dea80233d149c9 Message-Id: <20240410171856.91D213888C7F@sourceware.org> Date: Wed, 10 Apr 2024 17:18:56 +0000 (GMT) List-Id: https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=f55ba2fedc90438dcba7d3fdc4dea80233d149c9 commit f55ba2fedc90438dcba7d3fdc4dea80233d149c9 Author: Wilco Dijkstra Date: Wed Jan 11 13:51:17 2023 +0000 AArch64: Optimize memrchr Optimize the main loop - large strings are 43% faster on modern CPUs. Reviewed-by: Szabolcs Nagy (cherry picked from commit 00776241776e67fc666b896c1e85770f4f3ec1e1) Diff: --- sysdeps/aarch64/memrchr.S | 20 +++++++++++--------- 1 file changed, 11 insertions(+), 9 deletions(-) diff --git a/sysdeps/aarch64/memrchr.S b/sysdeps/aarch64/memrchr.S index 9925d8dcbf..e6a87f325e 100644 --- a/sysdeps/aarch64/memrchr.S +++ b/sysdeps/aarch64/memrchr.S @@ -26,7 +26,6 @@ * MTE compatible. */ -/* Arguments and results. */ #define srcin x0 #define chrin w1 #define cntin x2 @@ -77,31 +76,34 @@ ENTRY (__memrchr) csel result, result, xzr, hi ret + nop L(start_loop): - sub tmp, end, src - subs cntrem, cntin, tmp + subs cntrem, src, srcin b.ls L(nomatch) /* Make sure that it won't overread by a 16-byte chunk */ - add tmp, cntrem, 15 - tbnz tmp, 4, L(loop32_2) + sub cntrem, cntrem, 1 + tbz cntrem, 4, L(loop32_2) + add src, src, 16 - .p2align 4 + .p2align 5 L(loop32): - ldr qdata, [src, -16]! + ldr qdata, [src, -32]! cmeq vhas_chr.16b, vdata.16b, vrepchr.16b umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ fmov synd, dend cbnz synd, L(end) L(loop32_2): - ldr qdata, [src, -16]! + ldr qdata, [src, -16] subs cntrem, cntrem, 32 cmeq vhas_chr.16b, vdata.16b, vrepchr.16b - b.ls L(end) + b.lo L(end_2) umaxp vend.16b, vhas_chr.16b, vhas_chr.16b /* 128->64 */ fmov synd, dend cbz synd, L(loop32) +L(end_2): + sub src, src, 16 L(end): shrn vend.8b, vhas_chr.8h, 4 /* 128->64 */ fmov synd, dend