1/* 2 Copyright (c) 2024, Synopsys, Inc. All rights reserved. 3 4 Redistribution and use in source and binary forms, with or without 5 modification, are permitted provided that the following conditions are met: 6 7 1) Redistributions of source code must retain the above copyright notice, 8 this list of conditions and the following disclaimer. 9 10 2) Redistributions in binary form must reproduce the above copyright notice, 11 this list of conditions and the following disclaimer in the documentation 12 and/or other materials provided with the distribution. 13 14 3) Neither the name of the Synopsys, Inc., nor the names of its contributors 15 may be used to endorse or promote products derived from this software 16 without specific prior written permission. 17 18 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 19 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE 22 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 23 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 24 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 25 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 26 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 27 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 28 POSSIBILITY OF SUCH DAMAGE. 29*/ 30 31#include <picolibc.h> 32 33#include <sys/asm.h> 34 35#if defined (__ARC64_ARCH32__) 36 37; 64 bit version has the same working principles, with slightly different 38; instructions, so it is more commented 39 40ENTRY (strcmp) 41 xor r12, r12, r12 42 43 mov r8, NULL_32DT_1 44 45 asl r9, r8, 7 46 47.L_3_4B_comparison: 48 49 ld.ab r6, [r0, +4] 50 51 ld.ab r7, [r1, +4] 52 53#if defined (__ARC64_LL64__) 54 55 ldd.ab r2r3, [r0, +8] 56 57 ldd.ab r4r5, [r1, +8] 58 59#else 60 61 ld.ab r2, [r0, +4] 62 ld.ab r3, [r0, +4] 63 64 ld.ab r4, [r1, +4] 65 ld.ab r5, [r1, +4] 66 67#endif 68 69 sub r13, r6, r8 70 sub r10, r2, r8 71 sub r11, r3, r8 72 73 bic r13, r13, r6 74 bic r10, r10, r2 75 bic r11, r11, r3 76 77 ; Look for difference 78 sub.f 0, r6, r7 79 bset.ne r12, r12, 3 80 81 sub.f 0, r2, r4 82 bset.ne r12, r12, 2 83 84 sub.f 0, r3, r5 85 bset.ne r12, r12, 1 86 87 88 ; Look for NULL byte 89 and.f r13, r13, r9 90 bset.ne r12, r12, 3 91 92 and.f r10, r10, r9 93 bset.ne r12, r12, 2 94 95 and.f r11, r11, r9 96 bset.ne r12, r12, 1 97 98 breq r12, 0, @.L_3_4B_comparison 99 100; Setup r0, r3 and r5 with the relevant loaded and intermediate values 101 mov r0, r11 102 mov r3, r3 103 mov r5, r5 104 105 asr.f r12, r12, 3 106 107 mov.c r0, r10 108 mov.c r3, r2 109 mov.c r5, r4 110 111 asr.f r12, r12, 1 112 113 mov.c r0, r13 114 mov.c r3, r6 115 mov.c r5, r7 116 117 118 ffs.f r10, r0 119 xor r12, r3, r5 120 121 mov.z r10, 32 122 ffs r12, r12 123 124 xbfu r10, r10, 0b0111000011 125 xbfu r12, r12, 0b0111000011 126 127 128 sub.f 0, r10, r12 129 130 asl.ge r12, r12, 3 131 132; Difference is first 133 lsr.ge r3, r3, r12 134 lsr.ge r5, r5, r12 135 136 bmsk r3, r3, 7 137 bmsk r5, r5, 7 138 139 j_s.d [blink] 140 sub r0, r3, r5 141 142 143ENDFUNC(strcmp) 144 145#else 146 147ENTRY (strcmp) 148 149 xorl r12, r12, r12 150 151; Setup byte detector (more information bellow) [1] 152 vpack2wl r8, NULL_32DT_1, NULL_32DT_1 153; Set r9 as a copy of r8 for vectorized sub 154 asll r9, r8, 7 155 156.L_3_8B_comparison: 157 158 ldl.ab r6, [r0, +8] 159 160 ldl.ab r7, [r1, +8] 161 162; Using 128-bit memory operations 163#if defined (__ARC64_M128__) 164 165 lddl.ab r2r3, [r0, +16] 166 167 lddl.ab r4r5, [r1, +16] 168 169; The 64-bit crunching implementation. 170#elif defined (__ARC64_ARCH64__) 171 172 ldl.ab r2, [r0, +8] 173 ldl.ab r3, [r0, +8] 174 175 ldl.ab r4, [r1, +8] 176 ldl.ab r5, [r1, +8] 177 178#else 179 # error Unknown configuration 180#endif 181 182 subl r13, r6, r8 183 subl r10, r2, r8 184 subl r11, r3, r8 185 186 bicl r13, r13, r6 187 bicl r10, r10, r2 188 bicl r11, r11, r3 189 190; Look for difference 191 subl.f 0, r6, r7 192 bset.ne r12, r12, 3 193 194 subl.f 0, r2, r4 195 bset.ne r12, r12, 2 196 197 subl.f 0, r3, r5 198 bset.ne r12, r12, 1 199 200; Look for NULL byte 201 andl.f r13, r13, r9 202 bset.ne r12, r12, 3 203 204 andl.f r10, r10, r9 205 bset.ne r12, r12, 2 206 207 andl.f r11, r11, r9 208 bset.ne r12, r12, 1 209 210 breq r12, 0, @.L_3_8B_comparison 211 212; Setup r0, r3 and r5 with the relevant loaded and intermediate values [2] 213 ; [3] 214 movl r0, r11 215 movl r3, r3 216 movl r5, r5 217 218 asr.f r12, r12, 3 219 220 movl.c r0, r10 221 movl.c r3, r2 222 movl.c r5, r4 223 224 asr.f r12, r12, 1 225 226 movl.c r0, r13 227 movl.c r3, r6 228 movl.c r5, r7 229 230 ffsl.f r10, r0 ; [5] 231 xorl r12, r3, r5 232 233 movl.z r10, 64 ; [6] 234 ffsl r12, r12 ; [8] 235 236 xbful r10, r10, 0b0111000011 ; [7] 237 xbful r12, r12, 0b0111000011 238 239; r12 contains position of difference and r10 the position of a NULL byte 240; r3 and r5 contain the differing 8 bytes 241 242; Is there a difference? 243 subl.f 0, r10, r12 244; Multiply the byte position by 8 to get bit shift 245 asll.ge r12, r12, 3 246 247 lsrl.ge r3, r3, r12 248 lsrl.ge r5, r5, r12 249 250; There is no difference. Up until the NULL byte which must be 251 252 bmskl r3, r3, 7 253 bmskl r5, r5, 7 254 255 j_s.d [blink] 256 subl r0, r3, r5 257 258 259ENDFUNC (strcmp) 260 261#endif 262 263;; One important thing to note, is that we look for the first byte difference on 264;; both strings but we only look for the NULL byte in one string. 265;; This is because if a NULL byte appears first, it will be the first different 266;; byte. If it doesnt, the difference is what matters either way. If there is no 267;; difference, the NULL bytes will coincide! 268; 269; 270;; This code uses a common technique for NULL byte detection inside a word. 271;; Details on this technique can be found in: 272;; (https://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord) 273; 274; In sum, this technique allows for detecting a NULL byte inside any given 275; amount of bits by performing the following operation 276; DETECTNULL(X) (((X) - 0x01010101) & ~(X) & 0x80808080) [0] 277; 278; The code above implements this by setting r8 to a 0x01010101... sequence and 279; r1 to a 0x80808080... sequence of appropriate length 280; As LIMM are 32 bit only, we need to perform MOVHL and ORL [1] operations to 281; have the appropriate 64 bit values in place 282; 283;; Comparison is done 24 bytes at a time, either with 3 64 bit loads or 1 128 bit 284;; load and 1 64 bit. 285;; If either a NULL byte or a difference between the strings is found, r12 is 286;; used to know in which word the NULL/difference is found 287; 288; With the carry bit from r12, we can use mov.c to only move the appropriate 289; registers into the ones we will operate on [2]. We can safely directly move 290; the last set of registers without looking at r12, because if they aren't the 291; appropriate ones, they will be rewritten afterwards. [3] 292; 293;; Knowing the registers that contain the relevant information, we only need to 294;; look into where the difference and one of the zeros is. 295;; This is because, if the zeros are in different places, the difference will 296;; either be an earlier difference, or the first zero, so the actual zeros are 297;; irrelevant. 298;; Zero position is only relevant if there is no difference. And if there is no 299;; difference, the zeros have the same position. 300; 301; So now comes the tricky part. In order to obtain the position of a "first 302; NULL byte", we need to understand the NULL byte detection operation. 303; It is explained in depth in the link above but in short, it works by first 304; setting the highest bit of each byte to 1, if the corresponding byte is either 305; 0 or more than 0x80 306; Then, it makes the highest bit of each byte 1, if the byte is less than 0x80. 307; The last step is to AND these two values (this operation is simplified with 308; the SUB, BIC and TST instructions). 309; 310; This means that the evaluated equation result value has zeros for all non 311; zero bytes, except for the NULL bytes. Therefore, we can simply find the 312; first non zero bit (counting from bit 0) which will be inside the position of 313; the first NULL byte. [5] 314; 315; One thing to note, is that ffs oddly returns 31/63 if no bit is found, setting 316; the zero flag. As there can be that no NULL byte is present on one or both 317; strings at this point, we must set r10 and r11 to 32/64 when appropriate. [6] 318; 319; We can then convert the bit position into the last byte position by looking 320; into bits 3 to 5, and shifting 3 bits to the right. This can be combined into 321; a single xbful operation. The bottom 000011 represent shift by 3 and the top 322; 0111 represents the mask (3 to 5 shifted by 3 is 0 to 2). [7] 323; 324; To obtain the position of the difference, all we need to do is xor the two 325; registers. This way, every equal byte cancels out and all we are left with 326; is gibberish in the differing bytes. We can use the same ffs and xbuf 327; operations to get the differing byte position. 328; 329; Note that the order of the operations isnt the same as in this explanation, 330; to reduce register dependency between instructions 331; 332; 333; Unlike with r10, we dont need to check the zero flag for r12s' ffs because if 334; it is 0, it means there is no difference in the loaded data so any subtraction 335; operation will return 0 [8] 336; 337; There is one optimization that is being overlooked, which is returning 0 if 338; there is no difference, but there are NULL bytes anywhere, right after the 339; main loop. The reason for this is because the only way this can happen is if 340; the strings have the same length AND either are a multiple of 16/8 bytes, or 341; the bytes that follow the NULL bytes also match. As this is extremely 342; unlikely, it isnt worth it to perform this optimization since it would require 343; an extra branch in all runs 344; 345
