Home | History | Annotate | Line # | Download | only in string
memchr.S revision 1.3
      1  1.3  christos /*
      2  1.3  christos  * Written by J.T. Conklin <jtc (at) acorntoolworks.com>
      3  1.3  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       dsl 	RCSID("$NetBSD: memchr.S,v 1.3 2009/07/19 23:45:29 christos Exp $")
     10  1.1  christos #endif
     11  1.1  christos 
     12  1.3  christos ENTRY(memchr)
     13  1.3  christos 	movzbq	%sil,%rcx
     14  1.3  christos 
     15  1.3  christos 	/*
     16  1.3  christos 	 * Align to word boundary.
     17  1.3  christos 	 * Consider unrolling loop?
     18  1.3  christos 	 */
     19  1.3  christos 	testq	%rdx,%rdx	/* nbytes == 0? */
     20  1.3  christos 	je	.Lzero
     21  1.3  christos .Lalign:
     22  1.3  christos 	testb	$7,%dil
     23  1.3  christos 	je	.Lword_aligned
     24  1.3  christos 	movq	%rdi,%rax
     25  1.3  christos 	cmpb	(%rdi),%cl
     26  1.3  christos 	je	.Ldone
     27  1.3  christos 	incq	%rdi
     28  1.3  christos 	decq	%rdx
     29  1.3  christos 	jnz	.Lalign
     30  1.3  christos 	jmp	.Lzero
     31  1.3  christos 
     32  1.3  christos .Lword_aligned:
     33  1.3  christos 	/* copy char to all bytes in word */
     34  1.3  christos 	movb	%cl,%ch
     35  1.3  christos 	movq	%rcx,%rsi
     36  1.3  christos 	salq	$16,%rcx
     37  1.3  christos 	orq	%rsi,%rcx
     38  1.3  christos 	movq	%rcx,%rsi
     39  1.3  christos 	salq	$32,%rcx
     40  1.3  christos 	orq	%rsi,%rcx
     41  1.2       dsl 
     42  1.2       dsl 	movabsq	$0x0101010101010101,%r8
     43  1.3  christos 	movabsq	$0x8080808080808080,%r9
     44  1.1  christos 
     45  1.3  christos 	_ALIGN_TEXT
     46  1.3  christos .Lloop:
     47  1.3  christos 	cmpq	$7,%rdx		/* nbytes > 8 */
     48  1.3  christos 	jbe	.Lbyte
     49  1.3  christos 	movq	(%rdi),%rsi
     50  1.3  christos 	addq	$8,%rdi
     51  1.3  christos 	xorq	%rcx,%rsi
     52  1.3  christos 	subq	$8,%rdx
     53  1.3  christos 	subq	%r8,%rsi
     54  1.3  christos 	testq	%r9,%rsi
     55  1.3  christos 	je	.Lloop
     56  1.3  christos 
     57  1.3  christos 	/*
     58  1.3  christos 	 * In rare cases, the above loop may exit prematurely. We must
     59  1.3  christos 	 * return to the loop if none of the bytes in the word are
     60  1.3  christos 	 * equal to ch.
     61  1.3  christos 	 */
     62  1.3  christos 
     63  1.3  christos 	leaq	-8(%rdi),%rax
     64  1.3  christos 	cmpb	-8(%rdi),%cl	/* 1st byte == ch? */
     65  1.3  christos 	je	.Ldone
     66  1.3  christos 
     67  1.3  christos 	leaq	-7(%rdi),%rax
     68  1.3  christos 	cmpb	-7(%rdi),%cl	/* 2nd byte == ch? */
     69  1.3  christos 	je	.Ldone
     70  1.3  christos 
     71  1.3  christos 	leaq	-6(%rdi),%rax
     72  1.3  christos 	cmpb	-6(%rdi),%cl	/* 3rd byte == ch? */
     73  1.3  christos 	je	.Ldone
     74  1.3  christos 
     75  1.3  christos 	leaq	-5(%rdi),%rax
     76  1.3  christos 	cmpb	-5(%rdi),%cl	/* 4th byte == ch? */
     77  1.3  christos 	je	.Ldone
     78  1.3  christos 
     79  1.3  christos 	leaq	-4(%rdi),%rax
     80  1.3  christos 	cmpb	-4(%rdi),%cl	/* 5th byte == ch? */
     81  1.3  christos 	je	.Ldone
     82  1.3  christos 
     83  1.3  christos 	leaq	-3(%rdi),%rax
     84  1.3  christos 	cmpb	-3(%rdi),%cl	/* 6th byte == ch? */
     85  1.3  christos 	je	.Ldone
     86  1.3  christos 
     87  1.3  christos 	leaq	-2(%rdi),%rax
     88  1.3  christos 	cmpb	-2(%rdi),%cl	/* 7th byte == ch? */
     89  1.3  christos 	je	.Ldone
     90  1.3  christos 
     91  1.3  christos 	leaq	-1(%rdi),%rax
     92  1.3  christos 	cmpb	-1(%rdi),%cl	/* 7th byte == ch? */
     93  1.3  christos 	jne	.Lloop
     94  1.3  christos 	ret
     95  1.1  christos 
     96  1.3  christos .Lbyte:
     97  1.3  christos 	testq	%rdx,%rdx
     98  1.3  christos 	je	.Lzero
     99  1.3  christos .Lbyte_loop:
    100  1.3  christos 	movq	%rdi,%rax
    101  1.3  christos 	cmpb	(%rdi),%cl
    102  1.3  christos 	je	.Ldone
    103  1.3  christos 	incq	%rdi
    104  1.3  christos 	decq	%rdx
    105  1.3  christos 	jnz	.Lbyte_loop
    106  1.3  christos 
    107  1.3  christos .Lzero:
    108  1.3  christos 	xorq	%rax,%rax
    109  1.1  christos 
    110  1.3  christos .Ldone:
    111  1.1  christos 	ret
    112