Source-Changes-HG archive
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index][Old Index]
[src/trunk]: src/common/lib/libc/arch/arm/string Add a NEON(only) implementat...
details: https://anonhg.NetBSD.org/src/rev/a792abdf1aa8
branches: trunk
changeset: 783278:a792abdf1aa8
user: matt <matt%NetBSD.org@localhost>
date: Thu Dec 13 01:41:59 2012 +0000
description:
Add a NEON(only) implementation of memset.
This is a work in progress.
diffstat:
common/lib/libc/arch/arm/string/memset_neon.S | 186 ++++++++++++++++++++++++++
1 files changed, 186 insertions(+), 0 deletions(-)
diffs (190 lines):
diff -r bc05dbaa27be -r a792abdf1aa8 common/lib/libc/arch/arm/string/memset_neon.S
--- /dev/null Thu Jan 01 00:00:00 1970 +0000
+++ b/common/lib/libc/arch/arm/string/memset_neon.S Thu Dec 13 01:41:59 2012 +0000
@@ -0,0 +1,186 @@
+/* $NetBSD: memset_neon.S,v 1.1 2012/12/13 01:41:59 matt Exp $ */
+
+/*-
+ * Copyright (c) 2012 The NetBSD Foundation, Inc.
+ * All rights reserved.
+ *
+ * This code is derived from software contributed to The NetBSD Foundation
+ * by Matt Thomas of 3am Software Foundry.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
+ * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
+ * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+ * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
+ * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+ * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+ * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+ * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+ * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+ * POSSIBILITY OF SUCH DAMAGE.
+ */
+#include <machine/asm.h>
+#include "assym.h"
+
+/*
+ * memset: Sets a block of memory to the specified value
+ * Using NEON instructions
+ *
+ * On entry:
+ * r0 - dest address
+ * r1 - byte to write
+ * r2 - number of bytes to write
+ *
+ * On exit:
+ * r0 - dest address
+ */
+/* LINTSTUB: Func: void *memset(void *, int, size_t) */
+ENTRY(memset)
+ and r3, r1, #0xff /* We deal with bytes */
+ mov r1, r2
+ mov ip, r0 /* r0 needs to stay the same */
+
+ vdup.8 q0, r3 /* move fill to SIMD */
+ /* we no longer need to keep the fill value in a ARM register */
+
+ /* Ok first we will dword align the address */
+ ands r2, ip, #7 /* grab the bottom three bits */
+ beq .Lmemset_dwordaligned /* The addr is dword aligned */
+
+ bic ip, ip, #7 /* clear bottom three bits of addr */
+ vldr d7, [ip] /* load from memory */
+ add r1, r1, r2 /* add "pre-fill" to length */
+ lsl r2, r2, #3 /* byte to no-fill bit count */
+
+#ifdef __ARMEB__
+ neg r1, r1 /* start from the MSB */
+#endif
+ vmov s4, r2 /* move to SIMD d2 */
+ vmvn.u64 d3, #0 /* set all ones */
+ vshl.u64 d3, d3, d2 /* create a no-fill mask */
+ vmvn.u64 d2, d3 /* invert mask for a fill-mask */
+ vand d7, d7, d3 /* mask out fill bits */
+ vand d2, d0, d2 /* mask out no-fill bits */
+ vorr d7, d2, d7 /* merge fill and memory */
+
+ cmp r1, #8 /* Do we have less than 8 bytes */
+ movlt r2, #0 /* indicate this is the last word */
+ blt .Lmemset_lessthaneight_noload
+
+ vstmia ip!, {d7} /* write back to memory */
+ subs r1, r1, #8 /* and remove 8 bytes from the length */
+ RETc(eq)
+
+ /* We are now doubleword aligned */
+.Lmemset_dwordaligned:
+ vmov q1, q0 /* put fill in q1 (d2-d3) */
+ vmov q2, q0 /* put fill in q2 (d4-d5) */
+ vmov q3, q0 /* put fill in q3 (d6-d7) */
+
+ and r2, ip, #63 /* check for 64-byte alignment */
+ beq .Lmemset_8dwordaligned
+ /*
+ * Let's align to a 64-byte boundary so that stores don't cross
+ * cacheline boundaries. We also know we have at least 128-bytes to
+ * copy so we don't have to worry about the length at the moment.
+ */
+ rsb r2, r2, #64 /* how many bytes until 64 bytes */
+ cmp r1, r2 /* compare against length */
+ andlt r2, r1, #0x38 /* if < len, use trunc(len, 8) */
+ subs r1, r1, r2 /* subtract from len */
+ add pc, pc, r2 /* and jump to it */
+ nop
+ RETc(eq); b .Lmemset_lessthaneight
+ vstmia ip!, {d0}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d1}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d2}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d3}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d4}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d5}; b .Lmemset_8dwordaligned
+ vstmia ip!, {d0-d6}
+.Lmemset_8dwordaligned:
+ vmov d0, d1 /* restore in case of unaligned start */
+ cmp r1, #8 /* do we have less than 8 bytes */
+ movlt r2, #0 /* indicate last word */
+ blt .Lmemset_lessthaneight
+
+ cmp r1, #512
+ blt .Lmemset_sub512
+
+ /* Do 512 bytes at a time */
+ mov r2, #512
+.Lmemset_512:
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+ vstmia ip!, {d0-d7}
+.Lmemset_0:
+ subs r1, r1, r2
+ RETc(eq) /* return if done */
+ cmp r1, #512
+ bge .Lmemset_512
+
+ /*
+ * We have less than 512 bytes left, but since the sequence above
+ * store 64 bytes at a time, we determine the number of instructions
+ * we need to store the remainder (if >= 64 bytes) and execute that
+ * many vstmia.
+ */
+.Lmemset_sub512:
+ lsr r2, r1, #6 /* divide by 64 */
+ lslne r4, r2, #2 /* multiply by 4 */
+ addne r4, r4, #1f + 8 - .Lmemset_0
+ /* add the # of bytes between */
+1: subne pc, r4 /* and go */
+
+ /*
+ * We have less than 64 bytes to copy on a 8dword aligned address
+ */
+ and r2, r1, #56 /* get # of full dwords */
+ ands r1, r1, #7 /* get # of extra bytes */
+ beq .Lmemset_finalstore
+ /*
+ * The last word is a partial fill so load its value and update it
+ * to include the fill value.
+ */
+.Lmemset_lessthaneight:
+ vldr d7, [ip, r2] /* load the last partial dword */
+.Lmemset_lessthaneight_noload:
+ lsl r1, r1, #3 /* byte to fill bit count */
+#ifdef __ARMEB__
+ neg r1, r1 /* start from the MSB */
+#endif
+ vmov s4, r1 /* move to SIMD d2 */
+ vmvn.u64 d3, #0 /* set all ones */
+ vshl.u64 d3, d3, d2 /* create a no-fill mask */
+ vmvn.u64 d2, d3 /* invert mask */
+ vand d7, d7, d2 /* keep no-fill bits */
+ vand d2, d0, d3 /* mask out no-fill bits */
+ vorr d7, d2, d7 /* merge fill and no-fill */
+ vmov q1, q0 /* restore d2 & d3 */
+ add r2, r2, #8 /* compensate for the partial dword */
+.Lmemset_finalstore:
+ add pc, pc, r2 /* and jump to it */
+ nop
+ vstr d7, [ip]; RET
+ vstmia ip, {d6-d7}; RET
+ vstmia ip, {d5-d7}; RET
+ vstmia ip, {d4-d7}; RET
+ vstmia ip, {d3-d7}; RET
+ vstmia ip, {d2-d7}; RET
+ vstmia ip, {d1-d7}; RET
+ vstmia ip, {d0-d7}; RET
+END(memset)
Home |
Main Index |
Thread Index |
Old Index