1/* SH5 code Copyright 2002 SuperH Ltd. */ 2 3#include <picolibc.h> 4 5#include "asm.h" 6 7ENTRY(strcmp) 8 9#if __SHMEDIA__ 10 ld.ub r2,0,r4 11 pt/l quickret0,tr0 12 ld.ub r3,0,r5 13 ptabs r18,tr2 14 beqi/u r4,0,tr0 15 ld.ub r2,1,r6 16 bne/u r4,r5,tr0 17 pt/l quickret1,tr1 18 ld.ub r3,1,r7 19 beqi/u r6,0,tr1 20 ld.ub r2,2,r4 21 bne/u r6,r7,tr1 22 ld.ub r3,2,r5 23 beqi/u r4,0,tr0 24 ld.ub r2,3,r6 25 bne/u r4,r5,tr0 26 ld.ub r3,3,r7 27 beqi/u r6,0,tr1 28 ld.ub r2,4,r4 29 bne/u r6,r7,tr1 30 ld.ub r3,4,r5 31 beqi/u r4,0,tr0 32 ld.ub r2,5,r6 33 bne/u r4,r5,tr0 34 ld.ub r3,5,r7 35 beqi/u r6,0,tr1 36 ld.ub r2,6,r4 37 bne/u r6,r7,tr1 38 ld.ub r3,6,r5 39 beqi/u r4,0,tr0 40 ld.ub r2,7,r6 41 bne/u r4,r5,tr0 42 ld.ub r3,7,r7 43 beqi/u r6,0,tr1 44 sub r3,r2,r3 45 bne/u r6,r7,tr1 46 47 andi r2,-8,r2 48 add r3,r2,r3 49 ldlo.q r3,8,r23 50 pt r23_zero,tr0 51 shlli r3,3,r22 52 sub r63,r22,r20 53 movi 0x101,r6 54 mperm.w r6,r63,r6 55 SHLO r6,r22,r7 56 msubs.ub r7,r23,r8 57 pt loop,tr1 58 bnei/u r8,0,tr0 // r23_zero 59 pt found_zero,tr0 60 addi r3,15,r3 61 andi r3,-8,r3 62 sub r3,r2,r3 63 bne/l r7,r6,tr1 // loop 64 /* The strings are aligned to each other. */ 65 /* It is possible to have a loop with six cycles / iteration 66 by re-ordering the exit conditions, but then it needs extra 67 time and/or code to sort out the r4 != r5 case. */ 68 pt al_loop,tr1 69 pt al_found_zero,tr0 70al_loop: 71 ld.q r2,8,r4 72 ldx.q r2,r3,r5 73 addi r2,8,r2 74 mcmpeq.b r63,r4,r8 75 pt cmp_quad,tr3 76 bnei/u r8,0,tr0 // al_found_zero 77 beq/l r4,r5,tr1 // al_loop 78 blink tr3,r63 // cmp_quad 79 80 .balign 8 81quickret0: 82 sub r4,r5,r2 83 blink tr2,r63 84quickret1: 85 sub r6,r7,r2 86 blink tr2,r63 87 88loop: 89 ld.q r2,8,r4 90 ldx.q r2,r3,r19 91 addi r2,8,r2 92 msubs.ub r6,r4,r8 93 mcmpeq.b r63,r19,r9 94 SHHI r19,r20,r21 95 or r21,r23,r5 96 SHLO r19,r22,r23 97 bne/u r8,r9,tr0 // found_zero 98 beq/l r4,r5,tr1 // loop 99cmp_quad: 100#ifdef __LITTLE_ENDIAN__ 101 byterev r4,r4 102 byterev r5,r5 103#endif 104 cmpgtu r4,r5,r6 105 cmpgtu r5,r4,r7 106 sub r6,r7,r2 107 blink tr2,r63 108found_zero: 109 pt zero_now,tr0 110 pt cmp_quad,tr1 111 SHHI r9,r20,r7 112 bne/u r8,r7,tr0 // zero_now 113 bne/u r4,r5,tr1 // cmp_quad 114 SHLO r9,r22,r8 115r23_zero: 116 ld.q r2,8,r4 117 add r23,r63,r5 118zero_now: 119al_found_zero: 120/* We konw that one of the values has at lest one zero, and r8 holds 121 an 0x01 or 0xff mask for every zero found in one of the operands. 122 If both operands have the first zero in the same place, this mask 123 allows us to truncate the comparison to the valid bytes in the 124 strings. If the first zero is in different places, it doesn't 125 matter if some invalid bytes are included, since the comparison 126 of the zero with the non-zero will determine the outcome. */ 127#ifdef __LITTLE_ENDIAN__ 128 shlli r8,8,r8 129 addi r8,-1,r9 130 andc r9,r8,r8 131 and r8,r4,r4 132 and r8,r5,r5 133#else 134 shlri r8,1,r8 135 nsb r8,r8 136 addi r8,8,r8 137 andi r8,56,r8 138 sub r63,r8,r8 139 shlrd r4,r8,r4 140 shlrd r5,r8,r5 141#endif 142#ifdef __LITTLE_ENDIAN__ 143 byterev r4,r4 144 byterev r5,r5 145#endif 146 cmpgtu r4,r5,r6 147 cmpgtu r5,r4,r7 148 sub r6,r7,r2 149 blink tr2,r63 150 151#else /* ! __SHMEDIA__, i.e. SH 1..4 / SHcompact */ 152 153#ifdef __SH5__ 154#define STR1 r2 155#define STR2 r3 156#define RESULT r2 157#define TMP r4 158#else 159! Entry: r4: string1 160! r5: string2 161! Exit: r0: result 162! r1-r2,r4-r5: clobbered 163#define STR1 r4 164#define STR2 r5 165#define RESULT r0 166#define TMP r2 167#endif /* __SH5__ */ 168 169 mov STR1,r0 170 or STR2,r0 171 tst #3,r0 172 bf L_setup_char_loop 173 mov #0,r0 174#ifdef DELAYED_BRANCHES 175 mov.l @STR1+,r1 176 .align 2 177Longword_loop: 178 mov.l @STR2+,TMP 179 cmp/str r0,r1 180 bt Longword_loop_end 181 cmp/eq r1,TMP 182 bt.s Longword_loop 183 mov.l @STR1+,r1 184 add #-4, STR1 185Longword_loop_end: 186 add #-4, STR1 187 add #-4, STR2 188L_setup_char_loop: 189 mov.b @STR1+,r0 190 .align 2 191L_char_loop: 192 mov.b @STR2+,r1 193 tst r0,r0 194 bt L_return 195 cmp/eq r0,r1 196 bt.s L_char_loop 197 mov.b @STR1+,r0 198 add #-2,STR1 199 mov.b @STR1,r0 200#else /* ! DELAYED_BRANCHES */ 201 .align 2 202Longword_loop: 203 mov.l @r4+,r1 204 mov.l @r5+,r2 205 cmp/str r0,r1 206 bt Longword_loop_end 207 cmp/eq r1,r2 208 bt Longword_loop 209Longword_loop_end: 210 add #-4, r4 211 add #-4, r5 212 .align 2 213L_setup_char_loop: 214L_char_loop: 215 mov.b @r4+,r0 216 mov.b @r5+,r1 217 tst r0,r0 218 bt L_return 219 cmp/eq r0,r1 220 bt L_char_loop 221#endif 222L_return: 223 extu.b r0,RESULT 224 extu.b r1,r1 225 rts 226 sub r1,RESULT 227#endif /* ! __SHMEDIA__ */ 228
