memset_arm.S revision 1.1 1 1.1 matt /* $NetBSD: memset_arm.S,v 1.1 2013/01/12 20:27:13 matt Exp $ */
2 1.1 matt
3 1.1 matt /*-
4 1.1 matt * Copyright (c) 2012 The NetBSD Foundation, Inc.
5 1.1 matt * All rights reserved.
6 1.1 matt *
7 1.1 matt * This code is derived from software contributed to The NetBSD Foundation
8 1.1 matt * by Matt Thomas of 3am Software Foundry.
9 1.1 matt *
10 1.1 matt * Redistribution and use in source and binary forms, with or without
11 1.1 matt * modification, are permitted provided that the following conditions
12 1.1 matt * are met:
13 1.1 matt * 1. Redistributions of source code must retain the above copyright
14 1.1 matt * notice, this list of conditions and the following disclaimer.
15 1.1 matt * 2. Redistributions in binary form must reproduce the above copyright
16 1.1 matt * notice, this list of conditions and the following disclaimer in the
17 1.1 matt * documentation and/or other materials provided with the distribution.
18 1.1 matt *
19 1.1 matt * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 1.1 matt * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 1.1 matt * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 1.1 matt * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 1.1 matt * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 1.1 matt * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 1.1 matt * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 1.1 matt * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 1.1 matt * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 1.1 matt * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 1.1 matt * POSSIBILITY OF SUCH DAMAGE.
30 1.1 matt */
31 1.1 matt #include <machine/asm.h>
32 1.1 matt
33 1.1 matt #if defined(NEON)
34 1.1 matt #define STORE8 vst1.32 {d0}, [ip:64]!
35 1.1 matt #define STORE16 vst1.32 {d0-d1}, [ip:64]!
36 1.1 matt #define STORE32 vst1.32 {d0-d3}, [ip:64]!
37 1.1 matt #elif defined(VFP)
38 1.1 matt #define STORE8 vstmia ip!, {d0}
39 1.1 matt #define STORE16 vstmia ip!, {d0-d1}
40 1.1 matt #define STORE32 vstmia ip!, {d0-d3}
41 1.1 matt #elif defined(_ARM_ARCH_DWORD_OK)
42 1.1 matt #define STORE8 strd r2, [ip], #8
43 1.1 matt #define STORE16 STORE8; STORE8
44 1.1 matt #define STORE32 STORE16; STORE16
45 1.1 matt #else
46 1.1 matt #define STORE8 stmia ip!, {r2,r3}
47 1.1 matt #define STORE16 STORE8; STORE8
48 1.1 matt #define STORE32 STORE16; STORE16
49 1.1 matt #endif
50 1.1 matt /*
51 1.1 matt * memset: Sets a block of memory to the specified value
52 1.1 matt * Using NEON instructions
53 1.1 matt *
54 1.1 matt * On entry:
55 1.1 matt * r0 - dest address
56 1.1 matt * r1 - byte to write
57 1.1 matt * r2 - number of bytes to write
58 1.1 matt *
59 1.1 matt * On exit:
60 1.1 matt * r0 - dest address
61 1.1 matt */
62 1.1 matt /* LINTSTUB: Func: void *memset(void *, int, size_t) */
63 1.1 matt ENTRY(memset)
64 1.1 matt ands r3, r1, #0xff /* We deal with bytes */
65 1.1 matt orrne r3, r3, r3, lsl #8 /* replicate to all bytes */
66 1.1 matt orrne r3, r3, r3, lsl #16 /* replicate to all bytes */
67 1.1 matt movs r1, r2 /* we need r2 & r3 */
68 1.1 matt RETc(eq) /* return if length is 0 */
69 1.1 matt mov ip, r0 /* r0 needs to stay the same */
70 1.1 matt
71 1.1 matt cmp r1, #12 /* is this a small memset? *?
72 1.1 matt blt .Lbyte_by_byte /* then do it byte by byte */
73 1.1 matt
74 1.1 matt /* Ok first we will dword align the address */
75 1.1 matt ands r2, ip, #7 /* grab the bottom three bits */
76 1.1 matt beq .Lmemset_dwordaligned /* The addr is dword aligned */
77 1.1 matt
78 1.1 matt rsb r2, r2, #8 /* how far until dword aligned? */
79 1.1 matt sub r1, r1, r2 /* subtract it from remaining length */
80 1.1 matt mov r2, r3 /* duplicate fill value */
81 1.1 matt
82 1.1 matt tst ip, #1 /* halfword aligned? */
83 1.1 matt strneb r3, [ip], #1 /* no, write a byte */
84 1.1 matt tst ip, #2 /* word aligned? */
85 1.1 matt strneh r3, [ip], #2 /* no, write a halfword */
86 1.1 matt tst ip, #4 /* dword aligned? */
87 1.1 matt strne r3, [ip], #4 /* no, write a word */
88 1.1 matt
89 1.1 matt /* We are now doubleword aligned */
90 1.1 matt .Lmemset_dwordaligned:
91 1.1 matt #if defined(NEON)
92 1.1 matt vdup.8 q0, r3 /* move fill to SIMD */
93 1.1 matt vmov q1, q0 /* put fill in q1 (d2-d3) */
94 1.1 matt #elif defined(VFP)
95 1.1 matt mov r2, r3 /* duplicate fill value */
96 1.1 matt vmov d0, r2, r3 /* move to VFP */
97 1.1 matt vmov d1, r2, r3
98 1.1 matt vmov d2, r2, r3
99 1.1 matt vmov d3, r2, r3
100 1.1 matt #endif
101 1.1 matt
102 1.1 matt #if 1
103 1.1 matt cmp r1, #128
104 1.1 matt blt .Lmemset_mainloop
105 1.1 matt ands r2, ip, #63 /* check for 64-byte alignment */
106 1.1 matt beq .Lmemset_mainloop
107 1.1 matt /*
108 1.1 matt * Let's align to a 64-byte boundary so that stores don't cross
109 1.1 matt * cacheline boundaries. We also know we have at least 128-bytes to
110 1.1 matt * copy so we don't have to worry about the length at the moment.
111 1.1 matt */
112 1.1 matt rsb r2, r2, #64 /* how many bytes until 64 bytes */
113 1.1 matt sub r1, r1, r2 ?* subtract from remaining length */
114 1.1 matt #if !defined(NEON) && !defined(VFP)
115 1.1 matt mov r2, r3 /* put fill back in r2 */
116 1.1 matt #endif
117 1.1 matt
118 1.1 matt tst ip, #8 /* quadword aligned? */
119 1.1 matt beq 1f /* yes */
120 1.1 matt STORE8 /* no, store a dword */
121 1.1 matt 1: tst ip, #16 /* octaword aligned? *?
122 1.1 matt beq 2f /* yes */
123 1.1 matt STORE16 /* no, store a quadword */
124 1.1 matt 2: tst ip, #32 /* 32 word aligned? */
125 1.1 matt beq .Lmemset_mainloop /* yes */
126 1.1 matt STORE32 /* no, make 64-byte aligned
127 1.1 matt #endif
128 1.1 matt
129 1.1 matt .Lmemset_mainloop:
130 1.1 matt #if !defined(NEON) && !defined(VFP)
131 1.1 matt mov r2, r3 /* put fill back in r2 */
132 1.1 matt #endif
133 1.1 matt subs r1, r1, #64 /* subtract an initial 64 */
134 1.1 matt blt .Lmemset_lessthan_64bytes
135 1.1 matt
136 1.1 matt 3: STORE32 /* store first octaword */
137 1.1 matt STORE32 /* store second octaword */
138 1.1 matt RETc(eq) /* return if done */
139 1.1 matt subs r1, r1, #64 /* subtract another 64 */
140 1.1 matt bge 3b /* and do other if still >= 0 */
141 1.1 matt .Lmemset_lessthan_64bytes:
142 1.1 matt tst r1, #32 /* do we have 16 bytes left? */
143 1.1 matt beq .Lmemset_lessthan_32bytes
144 1.1 matt STORE32 /* yes, store an octaword */
145 1.1 matt bics r1, r1, #32 /* subtract 16 */
146 1.1 matt RETc(eq) /* return if length is 0 */
147 1.1 matt .Lmemset_lessthan_32bytes:
148 1.1 matt tst r1, #16 /* do we have 16 bytes left? */
149 1.1 matt beq .Lmemset_lessthan_16bytes
150 1.1 matt STORE16 /* yes, store a quadword */
151 1.1 matt bics r1, r1, #16 /* subtract 16 */
152 1.1 matt RETc(eq) /* return if length is 0 */
153 1.1 matt .Lmemset_lessthan_16bytes:
154 1.1 matt tst r1, #8 /* do we have 8 bytes left? */
155 1.1 matt beq .Lmemset_lessthan_8bytes/* no */
156 1.1 matt STORE8 /* yes, store a dword */
157 1.1 matt bics r1, r1, #8 /* subtract 8 */
158 1.1 matt RETc(eq) /* return if length is 0 */
159 1.1 matt .Lmemset_lessthan_8bytes:
160 1.1 matt tst r1, #4 /* do we have a word left? */
161 1.1 matt strne r2, [ip], #4 /* yes, so write one */
162 1.1 matt tst r1, #2 /* do we have a halfword left? */
163 1.1 matt strneh r2, [ip], #2 /* yes, so write one */
164 1.1 matt tst r1, #1 /* do we have a byte left? */
165 1.1 matt strneb r2, [ip], #1 /* yes, so write one */
166 1.1 matt RET /* return */
167 1.1 matt
168 1.1 matt .Lbyte_by_byte:
169 1.1 matt subs r1, r1, #1 /* can we write a byte? */
170 1.1 matt RETc(lt) /* no, we're done */
171 1.1 matt strb r3, [ip], #1 /* yes, so do it */
172 1.1 matt b .Lbyte_by_byte /* try next byte */
173 1.1 matt END(memset)
174