From 4499bb3e1b14c5121bcf7334d9b754d20768bb2b Mon Sep 17 00:00:00 2001 From: Marcus Shawcroft Date: Wed, 23 Jan 2013 16:32:33 +0000 Subject: [PATCH] AArch64: Adding optimized strnlen implementation. --- ports/ChangeLog.aarch64 | 5 + ports/sysdeps/aarch64/strnlen.S | 161 ++++++++++++++++++++++++++++++++ ports/sysdeps/aarch64/sysdep.h | 16 ++++ 3 files changed, 182 insertions(+) create mode 100644 ports/sysdeps/aarch64/strnlen.S diff --git a/ports/ChangeLog.aarch64 b/ports/ChangeLog.aarch64 index c7487f5b7d..ba4792fcb3 100644 --- a/ports/ChangeLog.aarch64 +++ b/ports/ChangeLog.aarch64 @@ -1,3 +1,8 @@ +2013-01-23 Marcus Shawcroft + + * sysdeps/aarch64/sysdep.h (ENTRY_ALIGN_AND_PAD): New. + * sysdeps/aarch64/strnlen.S: New file. + 2013-01-17 Marcus Shawcroft * sysdeps/aarch64/strlen.S: New file. diff --git a/ports/sysdeps/aarch64/strnlen.S b/ports/sysdeps/aarch64/strnlen.S new file mode 100644 index 0000000000..e582e8ad2d --- /dev/null +++ b/ports/sysdeps/aarch64/strnlen.S @@ -0,0 +1,161 @@ +/* strnlen - calculate the length of a string with limit. + + 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 + */ + +/* Arguments and results. */ +#define srcin x0 +#define len x0 +#define limit x1 + +/* Locals and temporaries. */ +#define src x2 +#define data1 x3 +#define data2 x4 +#define data2a x5 +#define has_nul1 x6 +#define has_nul2 x7 +#define tmp1 x8 +#define tmp2 x9 +#define tmp3 x10 +#define tmp4 x11 +#define zeroones x12 +#define pos x13 +#define limit_wd x14 + +#define REP8_01 0x0101010101010101 +#define REP8_7f 0x7f7f7f7f7f7f7f7f +#define REP8_80 0x8080808080808080 + +ENTRY_ALIGN_AND_PAD (__strnlen, 6, 9) + cbz limit, L(hit_limit) + mov zeroones, #REP8_01 + bic src, srcin, #15 + ands tmp1, srcin, #15 + b.ne L(misaligned) + /* Calculate the number of full and partial words -1. */ + sub limit_wd, limit, #1 /* Limit != 0, so no underflow. */ + lsr limit_wd, limit_wd, #4 /* Convert to Qwords. */ + + /* 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. */ + /* The inner loop deals with two Dwords at a time. This has a + slightly higher start-up cost, but we should win quite quickly, + especially on cores with a high number of issue slots per + cycle, as we get much better parallelism out of the operations. */ + + /* Start of critial section -- keep to one 64Byte cache line. */ +L(loop): + ldp data1, data2, [src], #16 +L(realigned): + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + sub tmp3, data2, zeroones + orr tmp4, data2, #REP8_7f + bic has_nul1, tmp1, tmp2 + bic has_nul2, tmp3, tmp4 + subs limit_wd, limit_wd, #1 + orr tmp1, has_nul1, has_nul2 + ccmp tmp1, #0, #0, pl /* NZCV = 0000 */ + b.eq L(loop) + /* End of critical section -- keep to one 64Byte cache line. */ + + orr tmp1, has_nul1, has_nul2 + cbz tmp1, L(hit_limit) /* No null in final Qword. */ + + /* We know there's a null in the final Qword. The easiest thing + to do now is work out the length of the string and return + MIN (len, limit). */ + + sub len, src, srcin + cbz has_nul1, L(nul_in_data2) +#ifdef __AARCH64EB__ + mov data2, data1 +#endif + sub len, len, #8 + mov has_nul2, has_nul1 +L(nul_in_data2): +#ifdef __AARCH64EB__ + /* For big-endian, carry propagation (if the final byte in the + string is 0x01) means we cannot use has_nul directly. The + easiest way to get the correct byte is to byte-swap the data + and calculate the syndrome a second time. */ + rev data2, data2 + sub tmp1, data2, zeroones + orr tmp2, data2, #REP8_7f + bic has_nul2, tmp1, tmp2 +#endif + sub len, len, #8 + rev has_nul2, has_nul2 + clz pos, has_nul2 + add len, len, pos, lsr #3 /* Bits to bytes. */ + cmp len, limit + csel len, len, limit, ls /* Return the lower value. */ + RET + +L(misaligned): + /* Deal with a partial first word. + We're doing two things in parallel here; + 1) Calculate the number of words (but avoiding overflow if + limit is near ULONG_MAX) - to do this we need to work out + limit + tmp1 - 1 as a 65-bit value before shifting it; + 2) Load and mask the initial data words - we force the bytes + before the ones we are interested in to 0xff - this ensures + early bytes will not hit any zero detection. */ + sub limit_wd, limit, #1 + neg tmp4, tmp1 + cmp tmp1, #8 + + and tmp3, limit_wd, #15 + lsr limit_wd, limit_wd, #4 + mov tmp2, #~0 + + ldp data1, data2, [src], #16 + lsl tmp4, tmp4, #3 /* Bytes beyond alignment -> bits. */ + add tmp3, tmp3, tmp1 + +#ifdef __AARCH64EB__ + /* Big-endian. Early bytes are at MSB. */ + lsl tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */ +#else + /* Little-endian. Early bytes are at LSB. */ + lsr tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */ +#endif + add limit_wd, limit_wd, tmp3, lsr #4 + + orr data1, data1, tmp2 + orr data2a, data2, tmp2 + + csinv data1, data1, xzr, le + csel data2, data2, data2a, le + b L(realigned) + +L(hit_limit): + mov len, limit + RET +END (__strnlen) +weak_alias (__strnlen, strnlen) +libc_hidden_def (strnlen) diff --git a/ports/sysdeps/aarch64/sysdep.h b/ports/sysdeps/aarch64/sysdep.h index 6b75ada148..9349471e36 100644 --- a/ports/sysdeps/aarch64/sysdep.h +++ b/ports/sysdeps/aarch64/sysdep.h @@ -42,6 +42,22 @@ cfi_startproc; \ CALL_MCOUNT +/* Define an entry point visible from C with a specified alignment and + pre-padding with NOPs. This can be used to ensure that a critical + loop within a function is cache line aligned. Note this version + does not adjust the padding if CALL_MCOUNT is defined. */ + +#define ENTRY_ALIGN_AND_PAD(name, align, padding) \ + .globl C_SYMBOL_NAME(name); \ + .type C_SYMBOL_NAME(name),%function; \ + .p2align align; \ + .rep padding; \ + nop; \ + .endr; \ + C_LABEL(name) \ + cfi_startproc; \ + CALL_MCOUNT + #undef END #define END(name) \ cfi_endproc; \