xref: /minix3/common/lib/libc/arch/i386/string/strlen.S (revision 0a6a1f1d05b60e214de2f05a7310ddd1f0e590e7)
1b6cbf720SGianluca Guida/*
2b6cbf720SGianluca Guida * Written by J.T. Conklin <jtc@acorntoolworks.com>
3b6cbf720SGianluca Guida * Public domain.
4b6cbf720SGianluca Guida */
5b6cbf720SGianluca Guida
6b6cbf720SGianluca Guida#include <machine/asm.h>
7b6cbf720SGianluca Guida
8b6cbf720SGianluca Guida#if defined(LIBC_SCCS)
9*0a6a1f1dSLionel Sambuc	RCSID("$NetBSD: strlen.S,v 1.2 2014/03/22 19:38:46 jakllsch Exp $")
10b6cbf720SGianluca Guida#endif
11b6cbf720SGianluca Guida
12b6cbf720SGianluca GuidaENTRY(strlen)
13b6cbf720SGianluca Guida	movl	4(%esp),%eax
14b6cbf720SGianluca Guida
15b6cbf720SGianluca Guida.Lalign:
16b6cbf720SGianluca Guida	/* Consider unrolling loop? */
17b6cbf720SGianluca Guida	testb	$3,%al
18b6cbf720SGianluca Guida	je	.Lword_aligned
19b6cbf720SGianluca Guida	cmpb	$0,(%eax)
20b6cbf720SGianluca Guida	je	.Ldone
21b6cbf720SGianluca Guida	incl	%eax
22b6cbf720SGianluca Guida	jmp	.Lalign
23b6cbf720SGianluca Guida
24b6cbf720SGianluca Guida	/*
25b6cbf720SGianluca Guida	 * There are many well known branch-free sequences which are used
26b6cbf720SGianluca Guida	 * for determining whether a zero-byte is contained within a word.
27b6cbf720SGianluca Guida	 * These sequences are generally much more efficent than loading
28b6cbf720SGianluca Guida	 * and comparing each byte individually.
29b6cbf720SGianluca Guida	 *
30b6cbf720SGianluca Guida	 * The expression [1,2]:
31b6cbf720SGianluca Guida	 *
32b6cbf720SGianluca Guida	 * (1)  ~(((x & 0x7f7f7f7f) + 0x7f7f7f7f) | (x | 0x7f7f7f7f))
33b6cbf720SGianluca Guida	 *
34b6cbf720SGianluca Guida	 * evaluates to a non-zero value if any of the bytes in the
35b6cbf720SGianluca Guida	 * original word is zero.
36b6cbf720SGianluca Guida	 *
37b6cbf720SGianluca Guida	 * It also has the useful property that bytes in the result word
38b6cbf720SGianluca Guida	 * that correspond to non-zero bytes in the original word have
39b6cbf720SGianluca Guida	 * the value 0x00, while bytes corresponding to zero bytes have
40b6cbf720SGianluca Guida	 * the value 0x80. This allows calculation of the first (and
41b6cbf720SGianluca Guida	 * last) occurrence of a zero byte within the word (useful for C's
42b6cbf720SGianluca Guida	 * str* primitives) by counting the number of leading (or
43b6cbf720SGianluca Guida	 * trailing) zeros and dividing the result by 8.  On machines
44b6cbf720SGianluca Guida	 * without (or with slow) clz() / ctz() instructions, testing
45b6cbf720SGianluca Guida	 * each byte in the result word for zero is necessary.
46b6cbf720SGianluca Guida	 *
47b6cbf720SGianluca Guida	 * This typically takes 4 instructions (5 on machines without
48b6cbf720SGianluca Guida	 * "not-or") not including those needed to load the constant.
49b6cbf720SGianluca Guida	 *
50b6cbf720SGianluca Guida	 *
51b6cbf720SGianluca Guida	 * The expression:
52b6cbf720SGianluca Guida	 *
53b6cbf720SGianluca Guida	 * (2)  ((x - 0x01010101) & ~x & 0x80808080)
54b6cbf720SGianluca Guida	 *
55b6cbf720SGianluca Guida	 * evaluates to a non-zero value if any of the bytes in the
56b6cbf720SGianluca Guida	 * original word is zero.
57b6cbf720SGianluca Guida	 *
58b6cbf720SGianluca Guida	 * On little endian machines, the first byte in the result word
59b6cbf720SGianluca Guida	 * that corresponds to a zero byte in the original byte is 0x80,
60b6cbf720SGianluca Guida	 * so clz() can be used as above.  On big endian machines, and
61b6cbf720SGianluca Guida	 * little endian machines without (or with a slow) clz() insn,
62b6cbf720SGianluca Guida	 * testing each byte in the original for zero is necessary.
63b6cbf720SGianluca Guida	 *
64b6cbf720SGianluca Guida	 * This typically takes 3 instructions (4 on machines without
65b6cbf720SGianluca Guida	 * "and with complement") not including those needed to load
66b6cbf720SGianluca Guida	 * constants.
67b6cbf720SGianluca Guida	 *
68b6cbf720SGianluca Guida	 *
69b6cbf720SGianluca Guida	 * The expression:
70b6cbf720SGianluca Guida	 *
71b6cbf720SGianluca Guida	 * (3)  ((x - 0x01010101) & 0x80808080)
72b6cbf720SGianluca Guida	 *
73b6cbf720SGianluca Guida	 * always evaluates to a non-zero value if any of the bytes in
74b6cbf720SGianluca Guida	 * the original word is zero.  However, in rare cases, it also
75b6cbf720SGianluca Guida	 * evaluates to a non-zero value when none of the bytes in the
76b6cbf720SGianluca Guida	 * original word is zero.
77b6cbf720SGianluca Guida	 *
78b6cbf720SGianluca Guida	 * To account for possible false positives, each byte of the
79b6cbf720SGianluca Guida	 * original word must be checked when the expression evaluates to
80b6cbf720SGianluca Guida	 * a non-zero value.  However, because it is simpler than those
81b6cbf720SGianluca Guida	 * presented above, code that uses it will be faster as long as
82b6cbf720SGianluca Guida	 * the rate of false positives is low.
83b6cbf720SGianluca Guida	 *
84b6cbf720SGianluca Guida	 * This is likely, because the the false positive can only occur
85b6cbf720SGianluca Guida	 * if the most siginificant bit of a byte within the word is set.
86b6cbf720SGianluca Guida	 * The expression will never fail for typical 7-bit ASCII strings.
87b6cbf720SGianluca Guida	 *
88b6cbf720SGianluca Guida	 * This typically takes 2 instructions not including those needed
89b6cbf720SGianluca Guida	 * to load constants.
90b6cbf720SGianluca Guida	 *
91b6cbf720SGianluca Guida	 *
92b6cbf720SGianluca Guida	 * [1] Henry S. Warren Jr., "Hacker's Delight", Addison-Westley 2003
93b6cbf720SGianluca Guida	 *
94b6cbf720SGianluca Guida	 * [2] International Business Machines, "The PowerPC Compiler Writer's
95b6cbf720SGianluca Guida	 *     Guide", Warthman Associates, 1996
96b6cbf720SGianluca Guida	 */
97b6cbf720SGianluca Guida
98b6cbf720SGianluca Guida	_ALIGN_TEXT
99b6cbf720SGianluca Guida.Lword_aligned:
100b6cbf720SGianluca Guida.Lloop:
101b6cbf720SGianluca Guida	movl	(%eax),%ecx
102b6cbf720SGianluca Guida	addl	$4,%eax
103b6cbf720SGianluca Guida	leal	-0x01010101(%ecx),%edx
104b6cbf720SGianluca Guida	testl	$0x80808080,%edx
105b6cbf720SGianluca Guida	je	.Lloop
106b6cbf720SGianluca Guida
107b6cbf720SGianluca Guida	/*
108b6cbf720SGianluca Guida	 * In rare cases, the above loop may exit prematurely. We must
109b6cbf720SGianluca Guida	 * return to the loop if none of the bytes in the word equal 0.
110b6cbf720SGianluca Guida	 */
111b6cbf720SGianluca Guida
112b6cbf720SGianluca Guida	/*
113b6cbf720SGianluca Guida	 * The optimal code for determining whether each byte is zero
114b6cbf720SGianluca Guida	 * differs by processor.  This space-optimized code should be
115b6cbf720SGianluca Guida	 * acceptable on all, especially since we don't expect it to
116b6cbf720SGianluca Guida	 * be run frequently,
117b6cbf720SGianluca Guida	 */
118b6cbf720SGianluca Guida
119b6cbf720SGianluca Guida	testb	%cl,%cl		/* 1st byte == 0? */
120b6cbf720SGianluca Guida	jne	1f
121b6cbf720SGianluca Guida	subl	$4,%eax
122b6cbf720SGianluca Guida	jmp	.Ldone
123b6cbf720SGianluca Guida
124b6cbf720SGianluca Guida1:	testb	%ch,%ch		/* 2nd byte == 0? */
125b6cbf720SGianluca Guida	jne	1f
126b6cbf720SGianluca Guida	subl	$3,%eax
127b6cbf720SGianluca Guida	jmp	.Ldone
128b6cbf720SGianluca Guida
129b6cbf720SGianluca Guida1:	shrl	$16,%ecx
130b6cbf720SGianluca Guida	testb	%cl,%cl		/* 3rd byte == 0? */
131b6cbf720SGianluca Guida	jne	1f
132b6cbf720SGianluca Guida	subl	$2,%eax
133b6cbf720SGianluca Guida	jmp	.Ldone
134b6cbf720SGianluca Guida
135b6cbf720SGianluca Guida1:	testb	%ch,%ch		/* 4th byte == 0? */
136b6cbf720SGianluca Guida	jne	.Lloop
137b6cbf720SGianluca Guida	decl	%eax
138b6cbf720SGianluca Guida
139b6cbf720SGianluca Guida.Ldone:
140b6cbf720SGianluca Guida	subl	4(%esp),%eax
141b6cbf720SGianluca Guida	ret
142*0a6a1f1dSLionel SambucEND(strlen)
143