strcat.S revision 1.2 1 1.1 christos /*
2 1.1 christos * Written by J.T. Conklin <jtc (at) acorntoolworks.com>
3 1.1 christos * Public domain.
4 1.1 christos */
5 1.1 christos
6 1.1 christos #include <machine/asm.h>
7 1.1 christos
8 1.1 christos #if defined(LIBC_SCCS)
9 1.2 jakllsch RCSID("$NetBSD: strcat.S,v 1.2 2014/03/22 19:16:34 jakllsch Exp $")
10 1.1 christos #endif
11 1.1 christos
12 1.1 christos ENTRY(strcat)
13 1.1 christos movq %rdi,%rax
14 1.1 christos movabsq $0x0101010101010101,%r8
15 1.1 christos movabsq $0x8080808080808080,%r9
16 1.1 christos
17 1.1 christos /*
18 1.1 christos * Align destination to word boundary.
19 1.1 christos * Consider unrolling loop?
20 1.1 christos */
21 1.1 christos .Lscan:
22 1.1 christos .Lscan_align:
23 1.1 christos testb $7,%dil
24 1.1 christos je .Lscan_aligned
25 1.1 christos cmpb $0,(%rdi)
26 1.1 christos je .Lcopy
27 1.1 christos incq %rdi
28 1.1 christos jmp .Lscan_align
29 1.1 christos
30 1.1 christos _ALIGN_TEXT
31 1.1 christos .Lscan_aligned:
32 1.1 christos .Lscan_loop:
33 1.1 christos movq (%rdi),%rdx
34 1.1 christos addq $8,%rdi
35 1.1 christos subq %r8,%rdx
36 1.1 christos testq %r9,%rdx
37 1.1 christos je .Lscan_loop
38 1.1 christos
39 1.1 christos /*
40 1.1 christos * In rare cases, the above loop may exit prematurely. We must
41 1.1 christos * return to the loop if none of the bytes in the word equal 0.
42 1.1 christos */
43 1.1 christos
44 1.1 christos cmpb $0,-8(%rdi) /* 1st byte == 0? */
45 1.1 christos jne 1f
46 1.1 christos subq $8,%rdi
47 1.1 christos jmp .Lcopy
48 1.1 christos
49 1.1 christos 1: cmpb $0,-7(%rdi) /* 2nd byte == 0? */
50 1.1 christos jne 1f
51 1.1 christos subq $7,%rdi
52 1.1 christos jmp .Lcopy
53 1.1 christos
54 1.1 christos 1: cmpb $0,-6(%rdi) /* 3rd byte == 0? */
55 1.1 christos jne 1f
56 1.1 christos subq $6,%rdi
57 1.1 christos jmp .Lcopy
58 1.1 christos
59 1.1 christos 1: cmpb $0,-5(%rdi) /* 4th byte == 0? */
60 1.1 christos jne 1f
61 1.1 christos subq $5,%rdi
62 1.1 christos jmp .Lcopy
63 1.1 christos
64 1.1 christos 1: cmpb $0,-4(%rdi) /* 5th byte == 0? */
65 1.1 christos jne 1f
66 1.1 christos subq $4,%rdi
67 1.1 christos jmp .Lcopy
68 1.1 christos
69 1.1 christos 1: cmpb $0,-3(%rdi) /* 6th byte == 0? */
70 1.1 christos jne 1f
71 1.1 christos subq $3,%rdi
72 1.1 christos jmp .Lcopy
73 1.1 christos
74 1.1 christos 1: cmpb $0,-2(%rdi) /* 7th byte == 0? */
75 1.1 christos jne 1f
76 1.1 christos subq $2,%rdi
77 1.1 christos jmp .Lcopy
78 1.1 christos
79 1.1 christos 1: cmpb $0,-1(%rdi) /* 8th byte == 0? */
80 1.1 christos jne .Lscan_loop
81 1.1 christos subq $1,%rdi
82 1.1 christos
83 1.1 christos /*
84 1.1 christos * Align source to a word boundary.
85 1.1 christos * Consider unrolling loop?
86 1.1 christos */
87 1.1 christos .Lcopy:
88 1.1 christos .Lcopy_align:
89 1.1 christos testb $7,%sil
90 1.1 christos je .Lcopy_aligned
91 1.1 christos movb (%rsi),%dl
92 1.1 christos incq %rsi
93 1.1 christos movb %dl,(%rdi)
94 1.1 christos incq %rdi
95 1.1 christos testb %dl,%dl
96 1.1 christos jne .Lcopy_align
97 1.1 christos ret
98 1.1 christos
99 1.1 christos _ALIGN_TEXT
100 1.1 christos .Lcopy_loop:
101 1.1 christos movq %rdx,(%rdi)
102 1.1 christos addq $8,%rdi
103 1.1 christos .Lcopy_aligned:
104 1.1 christos movq (%rsi),%rdx
105 1.1 christos movq %rdx,%rcx
106 1.1 christos addq $8,%rsi
107 1.1 christos subq %r8,%rcx
108 1.1 christos testq %r9,%rcx
109 1.1 christos je .Lcopy_loop
110 1.1 christos
111 1.1 christos /*
112 1.1 christos * In rare cases, the above loop may exit prematurely. We must
113 1.1 christos * return to the loop if none of the bytes in the word equal 0.
114 1.1 christos */
115 1.1 christos
116 1.1 christos movb %dl,(%rdi)
117 1.1 christos incq %rdi
118 1.1 christos testb %dl,%dl /* 1st byte == 0? */
119 1.1 christos je .Ldone
120 1.1 christos
121 1.1 christos shrq $8,%rdx
122 1.1 christos movb %dl,(%rdi)
123 1.1 christos incq %rdi
124 1.1 christos testb %dl,%dl /* 2nd byte == 0? */
125 1.1 christos je .Ldone
126 1.1 christos
127 1.1 christos shrq $8,%rdx
128 1.1 christos movb %dl,(%rdi)
129 1.1 christos incq %rdi
130 1.1 christos testb %dl,%dl /* 3rd byte == 0? */
131 1.1 christos je .Ldone
132 1.1 christos
133 1.1 christos shrq $8,%rdx
134 1.1 christos movb %dl,(%rdi)
135 1.1 christos incq %rdi
136 1.1 christos testb %dl,%dl /* 4th byte == 0? */
137 1.1 christos je .Ldone
138 1.1 christos
139 1.1 christos shrq $8,%rdx
140 1.1 christos movb %dl,(%rdi)
141 1.1 christos incq %rdi
142 1.1 christos testb %dl,%dl /* 5th byte == 0? */
143 1.1 christos je .Ldone
144 1.1 christos
145 1.1 christos shrq $8,%rdx
146 1.1 christos movb %dl,(%rdi)
147 1.1 christos incq %rdi
148 1.1 christos testb %dl,%dl /* 6th byte == 0? */
149 1.1 christos je .Ldone
150 1.1 christos
151 1.1 christos shrq $8,%rdx
152 1.1 christos movb %dl,(%rdi)
153 1.1 christos incq %rdi
154 1.1 christos testb %dl,%dl /* 7th byte == 0? */
155 1.1 christos je .Ldone
156 1.1 christos
157 1.1 christos shrq $8,%rdx
158 1.1 christos movb %dl,(%rdi)
159 1.1 christos incq %rdi
160 1.1 christos testb %dl,%dl /* 8th byte == 0? */
161 1.1 christos jne .Lcopy_aligned
162 1.1 christos
163 1.1 christos .Ldone:
164 1.1 christos ret
165 1.2 jakllsch END(strcat)
166