| memcmp.S (3e3aced3354bdc75a1c254f161a76a6e0d41b688) | memcmp.S (31489a9a2653e123121e8ca39b4be802013d2b50) |
|---|---|
| 1/* $NetBSD: memcmp.S,v 1.3 2003/10/14 07:51:45 scw Exp $ */ 2 3/* 4 * Copyright 2003 Wasabi Systems, Inc. 5 * All rights reserved. 6 * 7 * Written by Steve C. Woodford for Wasabi Systems, Inc. 8 * --- 63 unchanged lines hidden (view full) --- 72 cmp r2, #0x06 73 beq .Lmemcmp_6bytes 74#endif 75 mov r0, #0x00 76 77 /* Are both addresses aligned the same way? */ 78 cmp r2, #0x00 79 eornes r3, ip, r1 | 1/* $NetBSD: memcmp.S,v 1.3 2003/10/14 07:51:45 scw Exp $ */ 2 3/* 4 * Copyright 2003 Wasabi Systems, Inc. 5 * All rights reserved. 6 * 7 * Written by Steve C. Woodford for Wasabi Systems, Inc. 8 * --- 63 unchanged lines hidden (view full) --- 72 cmp r2, #0x06 73 beq .Lmemcmp_6bytes 74#endif 75 mov r0, #0x00 76 77 /* Are both addresses aligned the same way? */ 78 cmp r2, #0x00 79 eornes r3, ip, r1 |
| 80 moveq pc, lr /* len == 0, or same addresses! */ | 80 RETeq /* len == 0, or same addresses! */ |
| 81 tst r3, #0x03 82 subne r2, r2, #0x01 83 bne .Lmemcmp_bytewise2 /* Badly aligned. Do it the slow way */ 84 85 /* Word-align the addresses, if necessary */ 86 sub r3, r1, #0x05 87 ands r3, r3, #0x03 88 add r3, r3, r3, lsl #1 89 addne pc, pc, r3, lsl #3 90 nop 91 92 /* Compare up to 3 bytes */ 93 ldrb r0, [ip], #0x01 94 ldrb r3, [r1], #0x01 95 subs r0, r0, r3 | 81 tst r3, #0x03 82 subne r2, r2, #0x01 83 bne .Lmemcmp_bytewise2 /* Badly aligned. Do it the slow way */ 84 85 /* Word-align the addresses, if necessary */ 86 sub r3, r1, #0x05 87 ands r3, r3, #0x03 88 add r3, r3, r3, lsl #1 89 addne pc, pc, r3, lsl #3 90 nop 91 92 /* Compare up to 3 bytes */ 93 ldrb r0, [ip], #0x01 94 ldrb r3, [r1], #0x01 95 subs r0, r0, r3 |
| 96 movne pc, lr | 96 RETne |
| 97 subs r2, r2, #0x01 | 97 subs r2, r2, #0x01 |
| 98 moveq pc, lr | 98 RETeq |
| 99 100 /* Compare up to 2 bytes */ 101 ldrb r0, [ip], #0x01 102 ldrb r3, [r1], #0x01 103 subs r0, r0, r3 | 99 100 /* Compare up to 2 bytes */ 101 ldrb r0, [ip], #0x01 102 ldrb r3, [r1], #0x01 103 subs r0, r0, r3 |
| 104 movne pc, lr | 104 RETne |
| 105 subs r2, r2, #0x01 | 105 subs r2, r2, #0x01 |
| 106 moveq pc, lr | 106 RETeq |
| 107 108 /* Compare 1 byte */ 109 ldrb r0, [ip], #0x01 110 ldrb r3, [r1], #0x01 111 subs r0, r0, r3 | 107 108 /* Compare 1 byte */ 109 ldrb r0, [ip], #0x01 110 ldrb r3, [r1], #0x01 111 subs r0, r0, r3 |
| 112 movne pc, lr | 112 RETne |
| 113 subs r2, r2, #0x01 | 113 subs r2, r2, #0x01 |
| 114 moveq pc, lr | 114 RETeq |
| 115 116 /* Compare 4 bytes at a time, if possible */ 117 subs r2, r2, #0x04 118 bcc .Lmemcmp_bytewise 119.Lmemcmp_word_aligned: 120 ldr r0, [ip], #0x04 121 ldr r3, [r1], #0x04 122 subs r2, r2, #0x04 123 cmpcs r0, r3 124 beq .Lmemcmp_word_aligned 125 sub r0, r0, r3 126 127 /* Correct for extra subtraction, and check if done */ 128 adds r2, r2, #0x04 129 cmpeq r0, #0x00 /* If done, did all bytes match? */ | 115 116 /* Compare 4 bytes at a time, if possible */ 117 subs r2, r2, #0x04 118 bcc .Lmemcmp_bytewise 119.Lmemcmp_word_aligned: 120 ldr r0, [ip], #0x04 121 ldr r3, [r1], #0x04 122 subs r2, r2, #0x04 123 cmpcs r0, r3 124 beq .Lmemcmp_word_aligned 125 sub r0, r0, r3 126 127 /* Correct for extra subtraction, and check if done */ 128 adds r2, r2, #0x04 129 cmpeq r0, #0x00 /* If done, did all bytes match? */ |
| 130 moveq pc, lr /* Yup. Just return */ | 130 RETeq /* Yup. Just return */ |
| 131 132 /* Re-do the final word byte-wise */ 133 sub ip, ip, #0x04 134 sub r1, r1, #0x04 135 136.Lmemcmp_bytewise: 137 add r2, r2, #0x03 138.Lmemcmp_bytewise2: 139 ldrb r0, [ip], #0x01 140 ldrb r3, [r1], #0x01 141 subs r2, r2, #0x01 142 cmpcs r0, r3 143 beq .Lmemcmp_bytewise2 144 sub r0, r0, r3 | 131 132 /* Re-do the final word byte-wise */ 133 sub ip, ip, #0x04 134 sub r1, r1, #0x04 135 136.Lmemcmp_bytewise: 137 add r2, r2, #0x03 138.Lmemcmp_bytewise2: 139 ldrb r0, [ip], #0x01 140 ldrb r3, [r1], #0x01 141 subs r2, r2, #0x01 142 cmpcs r0, r3 143 beq .Lmemcmp_bytewise2 144 sub r0, r0, r3 |
| 145 mov pc, lr | 145 RET |
| 146 147#if defined(_KERNEL) && !defined(_STANDALONE) 148 /* 149 * 6 byte compares are very common, thanks to the network stack. 150 * This code is hand-scheduled to reduce the number of stalls for 151 * load results. Everything else being equal, this will be ~32% 152 * faster than a byte-wise memcmp. 153 */ 154 .align 5 155.Lmemcmp_6bytes: 156 ldrb r3, [r1, #0x00] /* r3 = b2#0 */ 157 ldrb r0, [ip, #0x00] /* r0 = b1#0 */ 158 ldrb r2, [r1, #0x01] /* r2 = b2#1 */ 159 subs r0, r0, r3 /* r0 = b1#0 - b2#0 */ 160 ldreqb r3, [ip, #0x01] /* r3 = b1#1 */ | 146 147#if defined(_KERNEL) && !defined(_STANDALONE) 148 /* 149 * 6 byte compares are very common, thanks to the network stack. 150 * This code is hand-scheduled to reduce the number of stalls for 151 * load results. Everything else being equal, this will be ~32% 152 * faster than a byte-wise memcmp. 153 */ 154 .align 5 155.Lmemcmp_6bytes: 156 ldrb r3, [r1, #0x00] /* r3 = b2#0 */ 157 ldrb r0, [ip, #0x00] /* r0 = b1#0 */ 158 ldrb r2, [r1, #0x01] /* r2 = b2#1 */ 159 subs r0, r0, r3 /* r0 = b1#0 - b2#0 */ 160 ldreqb r3, [ip, #0x01] /* r3 = b1#1 */ |
| 161 movne pc, lr /* Return if mismatch on #0 */ | 161 RETne /* Return if mismatch on #0 */ |
| 162 subs r0, r3, r2 /* r0 = b1#1 - b2#1 */ 163 ldreqb r3, [r1, #0x02] /* r3 = b2#2 */ 164 ldreqb r0, [ip, #0x02] /* r0 = b1#2 */ | 162 subs r0, r3, r2 /* r0 = b1#1 - b2#1 */ 163 ldreqb r3, [r1, #0x02] /* r3 = b2#2 */ 164 ldreqb r0, [ip, #0x02] /* r0 = b1#2 */ |
| 165 movne pc, lr /* Return if mismatch on #1 */ | 165 RETne /* Return if mismatch on #1 */ |
| 166 ldrb r2, [r1, #0x03] /* r2 = b2#3 */ 167 subs r0, r0, r3 /* r0 = b1#2 - b2#2 */ 168 ldreqb r3, [ip, #0x03] /* r3 = b1#3 */ | 166 ldrb r2, [r1, #0x03] /* r2 = b2#3 */ 167 subs r0, r0, r3 /* r0 = b1#2 - b2#2 */ 168 ldreqb r3, [ip, #0x03] /* r3 = b1#3 */ |
| 169 movne pc, lr /* Return if mismatch on #2 */ | 169 RETne /* Return if mismatch on #2 */ |
| 170 subs r0, r3, r2 /* r0 = b1#3 - b2#3 */ 171 ldreqb r3, [r1, #0x04] /* r3 = b2#4 */ 172 ldreqb r0, [ip, #0x04] /* r0 = b1#4 */ | 170 subs r0, r3, r2 /* r0 = b1#3 - b2#3 */ 171 ldreqb r3, [r1, #0x04] /* r3 = b2#4 */ 172 ldreqb r0, [ip, #0x04] /* r0 = b1#4 */ |
| 173 movne pc, lr /* Return if mismatch on #3 */ | 173 RETne /* Return if mismatch on #3 */ |
| 174 ldrb r2, [r1, #0x05] /* r2 = b2#5 */ 175 subs r0, r0, r3 /* r0 = b1#4 - b2#4 */ 176 ldreqb r3, [ip, #0x05] /* r3 = b1#5 */ | 174 ldrb r2, [r1, #0x05] /* r2 = b2#5 */ 175 subs r0, r0, r3 /* r0 = b1#4 - b2#4 */ 176 ldreqb r3, [ip, #0x05] /* r3 = b1#5 */ |
| 177 movne pc, lr /* Return if mismatch on #4 */ | 177 RETne /* Return if mismatch on #4 */ |
| 178 sub r0, r3, r2 /* r0 = b1#5 - b2#5 */ | 178 sub r0, r3, r2 /* r0 = b1#5 - b2#5 */ |
| 179 mov pc, lr | 179 RET |
| 180#endif | 180#endif |