xref: /freebsd/lib/libc/amd64/string/stpcpy.S (revision c697fb7f)
1/*
2 * Adapted by Guillaume Morin <guillaume@morinfr.org> from strcpy.S
3 * written by J.T. Conklin <jtc@acorntoolworks.com>
4 * Public domain.
5 */
6
7#include <machine/asm.h>
8__FBSDID("$FreeBSD$");
9
10/*
11 * This stpcpy implementation copies a byte at a time until the
12 * source pointer is aligned to a word boundary, it then copies by
13 * words until it finds a word containing a zero byte, and finally
14 * copies by bytes until the end of the string is reached.
15 *
16 * While this may result in unaligned stores if the source and
17 * destination pointers are unaligned with respect to each other,
18 * it is still faster than either byte copies or the overhead of
19 * an implementation suitable for machines with strict alignment
20 * requirements.
21 */
22
23	.globl	stpcpy,__stpcpy
24ENTRY(stpcpy)
25__stpcpy:
26	movabsq $0x0101010101010101,%r8
27	movabsq $0x8080808080808080,%r9
28
29	/*
30	 * Align source to a word boundary.
31	 * Consider unrolling loop?
32	 */
33.Lalign:
34	testb	$7,%sil
35	je	.Lword_aligned
36	movb	(%rsi),%dl
37	incq	%rsi
38	movb	%dl,(%rdi)
39	incq	%rdi
40	testb	%dl,%dl
41	jne	.Lalign
42	movq	%rdi,%rax
43	dec	%rax
44	ret
45
46	.p2align 4
47.Lloop:
48	movq	%rdx,(%rdi)
49	addq	$8,%rdi
50.Lword_aligned:
51	movq	(%rsi),%rdx
52	movq	%rdx,%rcx
53	addq	$8,%rsi
54	subq	%r8,%rcx
55	testq	%r9,%rcx
56	je	.Lloop
57
58	/*
59	 * In rare cases, the above loop may exit prematurely. We must
60	 * return to the loop if none of the bytes in the word equal 0.
61	 */
62
63	movb	%dl,(%rdi)
64	testb	%dl,%dl		/* 1st byte == 0? */
65	je	.Ldone
66	incq	%rdi
67
68	shrq	$8,%rdx
69	movb	%dl,(%rdi)
70	testb	%dl,%dl		/* 2nd byte == 0? */
71	je	.Ldone
72	incq	%rdi
73
74	shrq	$8,%rdx
75	movb	%dl,(%rdi)
76	testb	%dl,%dl		/* 3rd byte == 0? */
77	je	.Ldone
78	incq	%rdi
79
80	shrq	$8,%rdx
81	movb	%dl,(%rdi)
82	testb	%dl,%dl		/* 4th byte == 0? */
83	je	.Ldone
84	incq	%rdi
85
86	shrq	$8,%rdx
87	movb	%dl,(%rdi)
88	testb	%dl,%dl		/* 5th byte == 0? */
89	je	.Ldone
90	incq	%rdi
91
92	shrq	$8,%rdx
93	movb	%dl,(%rdi)
94	testb	%dl,%dl		/* 6th byte == 0? */
95	je	.Ldone
96	incq	%rdi
97
98	shrq	$8,%rdx
99	movb	%dl,(%rdi)
100	testb	%dl,%dl		/* 7th byte == 0? */
101	je	.Ldone
102	incq	%rdi
103
104	shrq	$8,%rdx
105	movb	%dl,(%rdi)
106	incq	%rdi
107	testb	%dl,%dl		/* 8th byte == 0? */
108	jne	.Lword_aligned
109	decq	%rdi
110
111.Ldone:
112	movq	%rdi,%rax
113	ret
114END(stpcpy)
115
116	.section .note.GNU-stack,"",%progbits
117