1/* 2 * Written by J.T. Conklin <jtc@acorntoolworks.com> 3 * Public domain. 4 * 5 * $NetBSD: strcat.S,v 1.4 2004/07/26 18:51:21 drochner Exp $ 6 * $FreeBSD: src/lib/libc/amd64/string/strcat.S,v 1.2 2008/11/02 01:10:54 peter Exp $ 7 */ 8 9#include <machine/asm.h> 10 11ENTRY(strcat) 12 movq %rdi,%rax 13 movabsq $0x0101010101010101,%r8 14 movabsq $0x8080808080808080,%r9 15 16 /* 17 * Align destination to word boundary. 18 * Consider unrolling loop? 19 */ 20.Lscan: 21.Lscan_align: 22 testb $7,%dil 23 je .Lscan_aligned 24 cmpb $0,(%rdi) 25 je .Lcopy 26 incq %rdi 27 jmp .Lscan_align 28 29 .align 4 30.Lscan_aligned: 31.Lscan_loop: 32 movq (%rdi),%rdx 33 addq $8,%rdi 34 subq %r8,%rdx 35 testq %r9,%rdx 36 je .Lscan_loop 37 38 /* 39 * In rare cases, the above loop may exit prematurely. We must 40 * return to the loop if none of the bytes in the word equal 0. 41 */ 42 43 cmpb $0,-8(%rdi) /* 1st byte == 0? */ 44 jne 1f 45 subq $8,%rdi 46 jmp .Lcopy 47 481: cmpb $0,-7(%rdi) /* 2nd byte == 0? */ 49 jne 1f 50 subq $7,%rdi 51 jmp .Lcopy 52 531: cmpb $0,-6(%rdi) /* 3rd byte == 0? */ 54 jne 1f 55 subq $6,%rdi 56 jmp .Lcopy 57 581: cmpb $0,-5(%rdi) /* 4th byte == 0? */ 59 jne 1f 60 subq $5,%rdi 61 jmp .Lcopy 62 631: cmpb $0,-4(%rdi) /* 5th byte == 0? */ 64 jne 1f 65 subq $4,%rdi 66 jmp .Lcopy 67 681: cmpb $0,-3(%rdi) /* 6th byte == 0? */ 69 jne 1f 70 subq $3,%rdi 71 jmp .Lcopy 72 731: cmpb $0,-2(%rdi) /* 7th byte == 0? */ 74 jne 1f 75 subq $2,%rdi 76 jmp .Lcopy 77 781: cmpb $0,-1(%rdi) /* 8th byte == 0? */ 79 jne .Lscan_loop 80 subq $1,%rdi 81 82 /* 83 * Align source to a word boundary. 84 * Consider unrolling loop? 85 */ 86.Lcopy: 87.Lcopy_align: 88 testb $7,%sil 89 je .Lcopy_aligned 90 movb (%rsi),%dl 91 incq %rsi 92 movb %dl,(%rdi) 93 incq %rdi 94 testb %dl,%dl 95 jne .Lcopy_align 96 ret 97 98 .align 4 99.Lcopy_loop: 100 movq %rdx,(%rdi) 101 addq $8,%rdi 102.Lcopy_aligned: 103 movq (%rsi),%rdx 104 movq %rdx,%rcx 105 addq $8,%rsi 106 subq %r8,%rcx 107 testq %r9,%rcx 108 je .Lcopy_loop 109 110 /* 111 * In rare cases, the above loop may exit prematurely. We must 112 * return to the loop if none of the bytes in the word equal 0. 113 */ 114 115 movb %dl,(%rdi) 116 incq %rdi 117 testb %dl,%dl /* 1st byte == 0? */ 118 je .Ldone 119 120 shrq $8,%rdx 121 movb %dl,(%rdi) 122 incq %rdi 123 testb %dl,%dl /* 2nd byte == 0? */ 124 je .Ldone 125 126 shrq $8,%rdx 127 movb %dl,(%rdi) 128 incq %rdi 129 testb %dl,%dl /* 3rd byte == 0? */ 130 je .Ldone 131 132 shrq $8,%rdx 133 movb %dl,(%rdi) 134 incq %rdi 135 testb %dl,%dl /* 4th byte == 0? */ 136 je .Ldone 137 138 shrq $8,%rdx 139 movb %dl,(%rdi) 140 incq %rdi 141 testb %dl,%dl /* 5th byte == 0? */ 142 je .Ldone 143 144 shrq $8,%rdx 145 movb %dl,(%rdi) 146 incq %rdi 147 testb %dl,%dl /* 6th byte == 0? */ 148 je .Ldone 149 150 shrq $8,%rdx 151 movb %dl,(%rdi) 152 incq %rdi 153 testb %dl,%dl /* 7th byte == 0? */ 154 je .Ldone 155 156 shrq $8,%rdx 157 movb %dl,(%rdi) 158 incq %rdi 159 testb %dl,%dl /* 8th byte == 0? */ 160 jne .Lcopy_aligned 161 162.Ldone: 163 ret 164END(strcat) 165 166 .section .note.GNU-stack,"",%progbits 167