1 /* SPDX-License-Identifier: GPL-2.0 */ <<
2 /* 1 /*
3 * linux/arch/alpha/lib/memset.S 2 * linux/arch/alpha/lib/memset.S
4 * 3 *
5 * This is an efficient (and small) implementa 4 * This is an efficient (and small) implementation of the C library "memset()"
6 * function for the alpha. 5 * function for the alpha.
7 * 6 *
8 * (C) Copyright 1996 Linus Torvalds 7 * (C) Copyright 1996 Linus Torvalds
9 * 8 *
10 * This routine is "moral-ware": you are free 9 * This routine is "moral-ware": you are free to use it any way you wish, and
11 * the only obligation I put on you is a moral 10 * the only obligation I put on you is a moral one: if you make any improvements
12 * to the routine, please send me your improve 11 * to the routine, please send me your improvements for me to use similarly.
13 * 12 *
14 * The scheduling comments are according to th 13 * The scheduling comments are according to the EV5 documentation (and done by
15 * hand, so they might well be incorrect, plea 14 * hand, so they might well be incorrect, please do tell me about it..)
16 */ 15 */
17 #include <linux/export.h> !! 16 #include <asm/export.h>
18 .set noat 17 .set noat
19 .set noreorder 18 .set noreorder
20 .text 19 .text
21 .globl memset 20 .globl memset
22 .globl __memset 21 .globl __memset
23 .globl ___memset 22 .globl ___memset
24 .globl __memset16 !! 23 .globl __memsetw
25 .globl __constant_c_memset 24 .globl __constant_c_memset
26 25
27 .ent ___memset 26 .ent ___memset
28 .align 5 27 .align 5
29 ___memset: 28 ___memset:
30 .frame $30,0,$26,0 29 .frame $30,0,$26,0
31 .prologue 0 30 .prologue 0
32 31
33 and $17,255,$1 /* E1 */ 32 and $17,255,$1 /* E1 */
34 insbl $17,1,$17 /* .. E0 */ 33 insbl $17,1,$17 /* .. E0 */
35 bis $17,$1,$17 /* E0 (p-c lat 34 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
36 sll $17,16,$1 /* E1 (p-c lat 35 sll $17,16,$1 /* E1 (p-c latency, next cycle) */
37 36
38 bis $17,$1,$17 /* E0 (p-c lat 37 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
39 sll $17,32,$1 /* E1 (p-c lat 38 sll $17,32,$1 /* E1 (p-c latency, next cycle) */
40 bis $17,$1,$17 /* E0 (p-c lat 39 bis $17,$1,$17 /* E0 (p-c latency, next cycle) */
41 ldq_u $31,0($30) /* .. E1 */ 40 ldq_u $31,0($30) /* .. E1 */
42 41
43 .align 5 42 .align 5
44 __constant_c_memset: 43 __constant_c_memset:
45 addq $18,$16,$6 /* E0 */ 44 addq $18,$16,$6 /* E0 */
46 bis $16,$16,$0 /* .. E1 */ 45 bis $16,$16,$0 /* .. E1 */
47 xor $16,$6,$1 /* E0 */ 46 xor $16,$6,$1 /* E0 */
48 ble $18,end /* .. E1 */ 47 ble $18,end /* .. E1 */
49 48
50 bic $1,7,$1 /* E0 */ 49 bic $1,7,$1 /* E0 */
51 beq $1,within_one_quad /* .. E1 (note 50 beq $1,within_one_quad /* .. E1 (note EV5 zero-latency forwarding) */
52 and $16,7,$3 /* E0 */ 51 and $16,7,$3 /* E0 */
53 beq $3,aligned /* .. E1 (note 52 beq $3,aligned /* .. E1 (note EV5 zero-latency forwarding) */
54 53
55 ldq_u $4,0($16) /* E0 */ 54 ldq_u $4,0($16) /* E0 */
56 bis $16,$16,$5 /* .. E1 */ 55 bis $16,$16,$5 /* .. E1 */
57 insql $17,$16,$2 /* E0 */ 56 insql $17,$16,$2 /* E0 */
58 subq $3,8,$3 /* .. E1 */ 57 subq $3,8,$3 /* .. E1 */
59 58
60 addq $18,$3,$18 /* E0 $18 is 59 addq $18,$3,$18 /* E0 $18 is new count ($3 is negative) */
61 mskql $4,$16,$4 /* .. E1 (and 60 mskql $4,$16,$4 /* .. E1 (and possible load stall) */
62 subq $16,$3,$16 /* E0 $16 is 61 subq $16,$3,$16 /* E0 $16 is new aligned destination */
63 bis $2,$4,$1 /* .. E1 */ 62 bis $2,$4,$1 /* .. E1 */
64 63
65 bis $31,$31,$31 /* E0 */ 64 bis $31,$31,$31 /* E0 */
66 ldq_u $31,0($30) /* .. E1 */ 65 ldq_u $31,0($30) /* .. E1 */
67 stq_u $1,0($5) /* E0 */ 66 stq_u $1,0($5) /* E0 */
68 bis $31,$31,$31 /* .. E1 */ 67 bis $31,$31,$31 /* .. E1 */
69 68
70 .align 4 69 .align 4
71 aligned: 70 aligned:
72 sra $18,3,$3 /* E0 */ 71 sra $18,3,$3 /* E0 */
73 and $18,7,$18 /* .. E1 */ 72 and $18,7,$18 /* .. E1 */
74 bis $16,$16,$5 /* E0 */ 73 bis $16,$16,$5 /* E0 */
75 beq $3,no_quad /* .. E1 */ 74 beq $3,no_quad /* .. E1 */
76 75
77 .align 3 76 .align 3
78 loop: 77 loop:
79 stq $17,0($5) /* E0 */ 78 stq $17,0($5) /* E0 */
80 subq $3,1,$3 /* .. E1 */ 79 subq $3,1,$3 /* .. E1 */
81 addq $5,8,$5 /* E0 */ 80 addq $5,8,$5 /* E0 */
82 bne $3,loop /* .. E1 */ 81 bne $3,loop /* .. E1 */
83 82
84 no_quad: 83 no_quad:
85 bis $31,$31,$31 /* E0 */ 84 bis $31,$31,$31 /* E0 */
86 beq $18,end /* .. E1 */ 85 beq $18,end /* .. E1 */
87 ldq $7,0($5) /* E0 */ 86 ldq $7,0($5) /* E0 */
88 mskqh $7,$6,$2 /* .. E1 (and 87 mskqh $7,$6,$2 /* .. E1 (and load stall) */
89 88
90 insqh $17,$6,$4 /* E0 */ 89 insqh $17,$6,$4 /* E0 */
91 bis $2,$4,$1 /* .. E1 */ 90 bis $2,$4,$1 /* .. E1 */
92 stq $1,0($5) /* E0 */ 91 stq $1,0($5) /* E0 */
93 ret $31,($26),1 /* .. E1 */ 92 ret $31,($26),1 /* .. E1 */
94 93
95 .align 3 94 .align 3
96 within_one_quad: 95 within_one_quad:
97 ldq_u $1,0($16) /* E0 */ 96 ldq_u $1,0($16) /* E0 */
98 insql $17,$16,$2 /* E1 */ 97 insql $17,$16,$2 /* E1 */
99 mskql $1,$16,$4 /* E0 (after l 98 mskql $1,$16,$4 /* E0 (after load stall) */
100 bis $2,$4,$2 /* E0 */ 99 bis $2,$4,$2 /* E0 */
101 100
102 mskql $2,$6,$4 /* E0 */ 101 mskql $2,$6,$4 /* E0 */
103 mskqh $1,$6,$2 /* .. E1 */ 102 mskqh $1,$6,$2 /* .. E1 */
104 bis $2,$4,$1 /* E0 */ 103 bis $2,$4,$1 /* E0 */
105 stq_u $1,0($16) /* E0 */ 104 stq_u $1,0($16) /* E0 */
106 105
107 end: 106 end:
108 ret $31,($26),1 /* E1 */ 107 ret $31,($26),1 /* E1 */
109 .end ___memset 108 .end ___memset
110 EXPORT_SYMBOL(___memset) 109 EXPORT_SYMBOL(___memset)
111 EXPORT_SYMBOL(__constant_c_memset) 110 EXPORT_SYMBOL(__constant_c_memset)
112 111
113 .align 5 112 .align 5
114 .ent __memset16 !! 113 .ent __memsetw
115 __memset16: !! 114 __memsetw:
116 .prologue 0 115 .prologue 0
117 116
118 inswl $17,0,$1 /* E0 */ 117 inswl $17,0,$1 /* E0 */
119 inswl $17,2,$2 /* E0 */ 118 inswl $17,2,$2 /* E0 */
120 inswl $17,4,$3 /* E0 */ 119 inswl $17,4,$3 /* E0 */
121 or $1,$2,$1 /* .. E1 */ 120 or $1,$2,$1 /* .. E1 */
122 inswl $17,6,$4 /* E0 */ 121 inswl $17,6,$4 /* E0 */
123 or $1,$3,$1 /* .. E1 */ 122 or $1,$3,$1 /* .. E1 */
124 or $1,$4,$17 /* E0 */ 123 or $1,$4,$17 /* E0 */
125 br __constant_c_memset /* .. E1 */ 124 br __constant_c_memset /* .. E1 */
126 125
127 .end __memset16 !! 126 .end __memsetw
128 EXPORT_SYMBOL(__memset16) !! 127 EXPORT_SYMBOL(__memsetw)
129 128
130 memset = ___memset 129 memset = ___memset
131 __memset = ___memset 130 __memset = ___memset
132 EXPORT_SYMBOL(memset) 131 EXPORT_SYMBOL(memset)
133 EXPORT_SYMBOL(__memset) 132 EXPORT_SYMBOL(__memset)
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.