From a0b1cd8869ffe7d0a602d31a3629c8690987be59 Mon Sep 17 00:00:00 2001 From: Marcus Shawcroft Date: Wed, 23 Jan 2013 16:33:14 +0000 Subject: [PATCH] AArch64: Adding optimized strncmp implementation. --- ports/ChangeLog.aarch64 | 4 + ports/sysdeps/aarch64/strncmp.S | 204 ++++++++++++++++++++++++++++++++ 2 files changed, 208 insertions(+) create mode 100644 ports/sysdeps/aarch64/strncmp.S diff --git a/ports/ChangeLog.aarch64 b/ports/ChangeLog.aarch64 index ba4792fcb3..5a42510cb6 100644 --- a/ports/ChangeLog.aarch64 +++ b/ports/ChangeLog.aarch64 @@ -1,3 +1,7 @@ +2013-01-23 Marcus Shawcroft + + * sysdeps/aarch64/strncmp.S: New file. + 2013-01-23 Marcus Shawcroft * sysdeps/aarch64/sysdep.h (ENTRY_ALIGN_AND_PAD): New. diff --git a/ports/sysdeps/aarch64/strncmp.S b/ports/sysdeps/aarch64/strncmp.S new file mode 100644 index 0000000000..341d5ced5f --- /dev/null +++ b/ports/sysdeps/aarch64/strncmp.S @@ -0,0 +1,204 @@ +/* Copyright (C) 2013 Free Software Foundation, Inc. + + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library. If not, see + . */ + +#include + +/* Assumptions: + * + * ARMv8-a, AArch64 + */ + +#define REP8_01 0x0101010101010101 +#define REP8_7f 0x7f7f7f7f7f7f7f7f +#define REP8_80 0x8080808080808080 + +/* Parameters and result. */ +#define src1 x0 +#define src2 x1 +#define limit x2 +#define result x0 + +/* Internal variables. */ +#define data1 x3 +#define data1w w3 +#define data2 x4 +#define data2w w4 +#define has_nul x5 +#define diff x6 +#define syndrome x7 +#define tmp1 x8 +#define tmp2 x9 +#define tmp3 x10 +#define zeroones x11 +#define pos x12 +#define limit_wd x13 +#define mask x14 +#define endloop x15 + +ENTRY_ALIGN_AND_PAD (strncmp, 6, 7) + cbz limit, L(ret0) + eor tmp1, src1, src2 + mov zeroones, #REP8_01 + tst tmp1, #7 + b.ne L(misaligned8) + ands tmp1, src1, #7 + b.ne L(mutual_align) + /* Calculate the number of full and partial words -1. */ + sub limit_wd, limit, #1 /* limit != 0, so no underflow. */ + lsr limit_wd, limit_wd, #3 /* Convert to Dwords. */ + + /* NUL detection works on the principle that (X - 1) & (~X) & 0x80 + (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and + can be done in parallel across the entire word. */ + /* Start of performance-critical section -- one 64B cache line. */ +L(loop_aligned): + ldr data1, [src1], #8 + ldr data2, [src2], #8 +L(start_realigned): + subs limit_wd, limit_wd, #1 + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + eor diff, data1, data2 /* Non-zero if differences found. */ + csinv endloop, diff, xzr, pl /* Last Dword or differences. */ + bics has_nul, tmp1, tmp2 /* Non-zero if NUL terminator. */ + ccmp endloop, #0, #0, eq + b.eq L(loop_aligned) + /* End of performance-critical section -- one 64B cache line. */ + + /* Not reached the limit, must have found the end or a diff. */ + tbz limit_wd, #63, L(not_limit) + + /* Limit % 8 == 0 => all bytes significant. */ + ands limit, limit, #7 + b.eq L(not_limit) + + lsl limit, limit, #3 /* Bits -> bytes. */ + mov mask, #~0 +#ifdef __AARCH64EB__ + lsr mask, mask, limit +#else + lsl mask, mask, limit +#endif + bic data1, data1, mask + bic data2, data2, mask + + /* Make sure that the NUL byte is marked in the syndrome. */ + orr has_nul, has_nul, mask + +L(not_limit): + orr syndrome, diff, has_nul + +#ifndef __AARCH64EB__ + rev syndrome, syndrome + rev data1, data1 + /* The MS-non-zero bit of the syndrome marks either the first bit + that is different, or the top bit of the first zero byte. + Shifting left now will bring the critical information into the + top bits. */ + clz pos, syndrome + rev data2, data2 + lsl data1, data1, pos + lsl data2, data2, pos + /* But we need to zero-extend (char is unsigned) the value and then + perform a signed 32-bit subtraction. */ + lsr data1, data1, #56 + sub result, data1, data2, lsr #56 + RET +#else + /* For big-endian we cannot use the trick with the syndrome value + as carry-propagation can corrupt the upper bits if the trailing + bytes in the string contain 0x01. */ + /* However, if there is no NUL byte in the dword, we can generate + the result directly. We can't just subtract the bytes as the + MSB might be significant. */ + cbnz has_nul, 1f + cmp data1, data2 + cset result, ne + cneg result, result, lo + RET +1: + /* Re-compute the NUL-byte detection, using a byte-reversed value. */ + rev tmp3, data1 + sub tmp1, tmp3, zeroones + orr tmp2, tmp3, #REP8_7f + bic has_nul, tmp1, tmp2 + rev has_nul, has_nul + orr syndrome, diff, has_nul + clz pos, syndrome + /* The MS-non-zero bit of the syndrome marks either the first bit + that is different, or the top bit of the first zero byte. + Shifting left now will bring the critical information into the + top bits. */ + lsl data1, data1, pos + lsl data2, data2, pos + /* But we need to zero-extend (char is unsigned) the value and then + perform a signed 32-bit subtraction. */ + lsr data1, data1, #56 + sub result, data1, data2, lsr #56 + RET +#endif + +L(mutual_align): + /* Sources are mutually aligned, but are not currently at an + alignment boundary. Round down the addresses and then mask off + the bytes that precede the start point. + We also need to adjust the limit calculations, but without + overflowing if the limit is near ULONG_MAX. */ + bic src1, src1, #7 + bic src2, src2, #7 + ldr data1, [src1], #8 + neg tmp3, tmp1, lsl #3 /* 64 - bits(bytes beyond align). */ + ldr data2, [src2], #8 + mov tmp2, #~0 + sub limit_wd, limit, #1 /* limit != 0, so no underflow. */ +#ifdef __AARCH64EB__ + /* Big-endian. Early bytes are at MSB. */ + lsl tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */ +#else + /* Little-endian. Early bytes are at LSB. */ + lsr tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */ +#endif + and tmp3, limit_wd, #7 + lsr limit_wd, limit_wd, #3 + /* Adjust the limit. Only low 3 bits used, so overflow irrelevant. */ + add limit, limit, tmp1 + add tmp3, tmp3, tmp1 + orr data1, data1, tmp2 + orr data2, data2, tmp2 + add limit_wd, limit_wd, tmp3, lsr #3 + b L(start_realigned) + +L(ret0): + mov result, #0 + RET + + .p2align 6 +L(misaligned8): + sub limit, limit, #1 +1: + /* Perhaps we can do better than this. */ + ldrb data1w, [src1], #1 + ldrb data2w, [src2], #1 + subs limit, limit, #1 + ccmp data1w, #1, #0, cs /* NZCV = 0b0000. */ + ccmp data1w, data2w, #0, cs /* NZCV = 0b0000. */ + b.eq 1b + sub result, data1, data2 + RET +END (strncmp) +libc_hidden_builtin_def (strncmp)