From patchwork Wed Jan 7 09:58:22 2015 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Richard Earnshaw X-Patchwork-Id: 4540 Received: (qmail 12899 invoked by alias); 7 Jan 2015 09:58:30 -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 12888 invoked by uid 89); 7 Jan 2015 09:58:29 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=-1.0 required=5.0 tests=AWL, BAYES_00, SPF_PASS autolearn=ham version=3.3.2 X-HELO: service87.mimecast.com Message-ID: <54AD033E.509@arm.com> Date: Wed, 07 Jan 2015 09:58:22 +0000 From: Richard Earnshaw User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.3.0 MIME-Version: 1.0 To: Glibc Development List Subject: Re: [Patch, AArch64] Optimized strcpy References: <54917329.4090601@arm.com> <5491759B.4020704@arm.com> In-Reply-To: <5491759B.4020704@arm.com> X-MC-Unique: 115010709582402901 On 17/12/14 12:22, Richard Earnshaw wrote: > On 17/12/14 12:12, Richard Earnshaw wrote: >> This patch contains an optimized implementation of strcpy for AArch64 >> systems. Benchmarking shows that it is approximately 20-25% faster than >> the generic implementation across the board. >> >> R. >> >> Richard Earnshaw >> >> * sysdeps/aarch64/strcpy.S: New file. >> >> > Following the various discussions about the above, I've done some further tweaking of the code and indeed there some further performance improvements, particularly for short strings. I think this is likely to be the final version (at least, for 2.21). Changes this time around: - Add the ability to build the code as stpcpy(). - Small change to the page crossing check, that uses the same number of instructions, but could be faster on some micro-architectures. - For the slow (page crossing) check, once a page cross is known to occur, jump to the normal entry point. - For big-endian only, on the first check we pre-reverse the bytes so that we don't have to recalculate the syndrome in the (likely) case that the string is short. - For the initial unaligned fetch, detect zeros in the first and second DWords independently and jump to the relevant epilogue sequence directly. This eliminates another level of branching later on for the special cases when we have to use sub-dword sized stores - Other changes are mostly re-ordering of the hunks of code and micro-optimizations that fall out of the above changes. OK? * sysdeps/aarch64/strcpy.S: New file. * sysdeps/aarch64/stpcpy.S: New file. diff --git a/sysdeps/aarch64/stpcpy.S b/sysdeps/aarch64/stpcpy.S new file mode 100644 index 0000000..6576a75 --- /dev/null +++ b/sysdeps/aarch64/stpcpy.S @@ -0,0 +1,20 @@ +/* stpcpy - copy a string returning pointer to end. + Copyright (C) 2015 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 + . */ + +#define BUILD_STPCPY +#include "strcpy.S" diff --git a/sysdeps/aarch64/strcpy.S b/sysdeps/aarch64/strcpy.S new file mode 100644 index 0000000..28846fb --- /dev/null +++ b/sysdeps/aarch64/strcpy.S @@ -0,0 +1,326 @@ +/* strcpy/stpcpy - copy a string returning pointer to start/end. + Copyright (C) 2013-2015 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 + . */ + +/* To build as stpcpy, define BUILD_STPCPY before compiling this file. + + To test the page crossing code path more thoroughly, compile with + -DSTRCPY_TEST_PAGE_CROSS - this will force all unaligned copies through + the slower entry path. This option is not intended for production use. */ + +#include + +/* Assumptions: + * + * ARMv8-a, AArch64, unaligned accesses, min page size 4k. + */ + +/* Arguments and results. */ +#define dstin x0 +#define srcin x1 + +/* Locals and temporaries. */ +#define src x2 +#define dst x3 +#define data1 x4 +#define data1w w4 +#define data2 x5 +#define data2w w5 +#define has_nul1 x6 +#define has_nul2 x7 +#define tmp1 x8 +#define tmp2 x9 +#define tmp3 x10 +#define tmp4 x11 +#define zeroones x12 +#define data1a x13 +#define data2a x14 +#define pos x15 +#define len x16 +#define to_align x17 + +#ifdef BUILD_STPCPY +#define STRCPY __stpcpy +#else +#define STRCPY strcpy +#endif + + /* 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. */ + +#define REP8_01 0x0101010101010101 +#define REP8_7f 0x7f7f7f7f7f7f7f7f +#define REP8_80 0x8080808080808080 + + /* AArch64 systems have a minimum page size of 4k. We can do a quick + page size check for crossing this boundary on entry and if we + do not, then we can short-circuit much of the entry code. We + expect early page-crossing strings to be rare (probability of + 16/MIN_PAGE_SIZE ~= 0.4%), so the branch should be quite + predictable, even with random strings. + + We don't bother checking for larger page sizes, the cost of setting + up the correct page size is just not worth the extra gain from + a small reduction in the cases taking the slow path. Note that + we only care about whether the first fetch, which may be + misaligned, crosses a page boundary - after that we move to aligned + fetches for the remainder of the string. */ + +#ifdef STRCPY_TEST_PAGE_CROSS + /* Make everything that isn't Qword aligned look like a page cross. */ +#define MIN_PAGE_P2 4 +#else +#define MIN_PAGE_P2 12 +#endif + +#define MIN_PAGE_SIZE (1 << MIN_PAGE_P2) + +ENTRY_ALIGN (STRCPY, 6) + /* For moderately short strings, the fastest way to do the copy is to + calculate the length of the string in the same way as strlen, then + essentially do a memcpy of the result. This avoids the need for + multiple byte copies and further means that by the time we + reach the bulk copy loop we know we can always use DWord + accesses. We expect strcpy to rarely be called repeatedly + with the same source string, so branch prediction is likely to + always be difficult - we mitigate against this by preferring + conditional select operations over branches whenever this is + feasible. */ + and tmp2, srcin, #(MIN_PAGE_SIZE - 1) + mov zeroones, #REP8_01 + and to_align, srcin, #15 + cmp tmp2, #(MIN_PAGE_SIZE - 16) + neg tmp1, to_align + /* The first fetch will straddle a (possible) page boundary iff + srcin + 15 causes bit[MIN_PAGE_P2] to change value. A 16-byte + aligned string will never fail the page align check, so will + always take the fast path. */ + b.gt L(page_cross) + +L(page_cross_ok): + ldp data1, data2, [srcin] +#ifdef __AARCH64EB__ + /* Because we expect the end to be found within 16 characters + (profiling shows this is the most common case), it's worth + swapping the bytes now to save having to recalculate the + termination syndrome later. We preserve data1 and data2 + so that we can re-use the values later on. */ + rev tmp2, data1 + sub tmp1, tmp2, zeroones + orr tmp2, tmp2, #REP8_7f + bics has_nul1, tmp1, tmp2 + b.ne L(fp_le8) + rev tmp4, data2 + sub tmp3, tmp4, zeroones + orr tmp4, tmp4, #REP8_7f +#else + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + bics has_nul1, tmp1, tmp2 + b.ne L(fp_le8) + sub tmp3, data2, zeroones + orr tmp4, data2, #REP8_7f +#endif + bics has_nul2, tmp3, tmp4 + b.eq L(bulk_entry) + + /* The string is short (<=16 bytes). We don't know exactly how + short though, yet. Work out the exact length so that we can + quickly select the optimal copy strategy. */ +L(fp_gt8): + rev has_nul2, has_nul2 + clz pos, has_nul2 + mov tmp2, #56 + add dst, dstin, pos, lsr #3 /* Bits to bytes. */ + sub pos, tmp2, pos +#ifdef __AARCH64EB__ + lsr data2, data2, pos +#else + lsl data2, data2, pos +#endif + str data2, [dst, #1] + str data1, [dstin] +#ifdef BUILD_STPCPY + add dstin, dst, #8 +#endif + ret + +L(fp_le8): + rev has_nul1, has_nul1 + clz pos, has_nul1 + add dst, dstin, pos, lsr #3 /* Bits to bytes. */ + subs tmp2, pos, #24 /* Pos in bits. */ + b.lt L(fp_lt4) +#ifdef __AARCH64EB__ + mov tmp2, #56 + sub pos, tmp2, pos + lsr data2, data1, pos + lsr data1, data1, #32 +#else + lsr data2, data1, tmp2 +#endif + /* 4->7 bytes to copy. */ + str data2w, [dst, #-3] + str data1w, [dstin] +#ifdef BUILD_STPCPY + mov dstin, dst +#endif + ret +L(fp_lt4): + cbz pos, L(fp_lt2) + /* 2->3 bytes to copy. */ +#ifdef __AARCH64EB__ + lsr data1, data1, #48 +#endif + strh data1w, [dstin] + /* Fall-through, one byte (max) to go. */ +L(fp_lt2): + /* Null-terminated string. Last character must be zero! */ + strb wzr, [dst] +#ifdef BUILD_STPCPY + mov dstin, dst +#endif + ret + + .p2align 6 + /* Aligning here ensures that the entry code and main loop all lies + within one 64-byte cache line. */ +L(bulk_entry): + sub to_align, to_align, #16 + stp data1, data2, [dstin] + sub src, srcin, to_align + sub dst, dstin, to_align + b L(entry_no_page_cross) + + /* 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. */ +L(main_loop): + stp data1, data2, [dst], #16 +L(entry_no_page_cross): + ldp data1, data2, [src], #16 + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + sub tmp3, data2, zeroones + orr tmp4, data2, #REP8_7f + bic has_nul1, tmp1, tmp2 + bics has_nul2, tmp3, tmp4 + ccmp has_nul1, #0, #0, eq /* NZCV = 0000 */ + b.eq L(main_loop) + + /* Since we know we are copying at least 16 bytes, the fastest way + to deal with the tail is to determine the location of the + trailing NUL, then (re)copy the 16 bytes leading up to that. */ + cmp has_nul1, #0 +#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. */ + csel data1, data1, data2, ne + rev data1, data1 + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + bic has_nul1, tmp1, tmp2 +#else + csel has_nul1, has_nul1, has_nul2, ne +#endif + rev has_nul1, has_nul1 + clz pos, has_nul1 + add tmp1, pos, #72 + add pos, pos, #8 + csel pos, pos, tmp1, ne + add src, src, pos, lsr #3 + add dst, dst, pos, lsr #3 + ldp data1, data2, [src, #-32] + stp data1, data2, [dst, #-16] +#ifdef BUILD_STPCPY + sub dstin, dst, #1 +#endif + ret + +L(page_cross): + bic src, srcin, #15 + /* Start by loading two words at [srcin & ~15], then forcing the + bytes that precede srcin to 0xff. This means they never look + like termination bytes. */ + ldp data1, data2, [src] + lsl tmp1, tmp1, #3 /* Bytes beyond alignment -> bits. */ + tst to_align, #7 + csetm tmp2, ne +#ifdef __AARCH64EB__ + lsl tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */ +#else + lsr tmp2, tmp2, tmp1 /* Shift (tmp1 & 63). */ +#endif + orr data1, data1, tmp2 + orr data2a, data2, tmp2 + cmp to_align, #8 + csinv data1, data1, xzr, lt + csel data2, data2, data2a, lt + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + sub tmp3, data2, zeroones + orr tmp4, data2, #REP8_7f + bic has_nul1, tmp1, tmp2 + bics has_nul2, tmp3, tmp4 + ccmp has_nul1, #0, #0, eq /* NZCV = 0000 */ + b.eq L(page_cross_ok) + /* We now need to make data1 and data2 look like they've been + loaded directly from srcin. Do a rotate on the 128-bit value. */ + lsl tmp1, to_align, #3 /* Bytes->bits. */ + neg tmp2, to_align, lsl #3 +#ifdef __AARCH64EB__ + lsl data1a, data1, tmp1 + lsr tmp4, data2, tmp2 + lsl data2, data2, tmp1 + orr tmp4, tmp4, data1a + cmp to_align, #8 + csel data1, tmp4, data2, lt + rev tmp2, data1 + rev tmp4, data2 + sub tmp1, tmp2, zeroones + orr tmp2, tmp2, #REP8_7f + sub tmp3, tmp4, zeroones + orr tmp4, tmp4, #REP8_7f +#else + lsr data1a, data1, tmp1 + lsl tmp4, data2, tmp2 + lsr data2, data2, tmp1 + orr tmp4, tmp4, data1a + cmp to_align, #8 + csel data1, tmp4, data2, lt + sub tmp1, data1, zeroones + orr tmp2, data1, #REP8_7f + sub tmp3, data2, zeroones + orr tmp4, data2, #REP8_7f +#endif + bic has_nul1, tmp1, tmp2 + cbnz has_nul1, L(fp_le8) + bic has_nul2, tmp3, tmp4 + b L(fp_gt8) +END (STRCPY) + +#ifdef BUILD_STPCPY +weak_alias (__stpcpy, stpcpy) +libc_hidden_def (__stpcpy) +libc_hidden_builtin_def (stpcpy) +#else +libc_hidden_builtin_def (strcpy) +#endif