1*5d9d9091SRichard Lowe/* 2*5d9d9091SRichard Lowe * CDDL HEADER START 3*5d9d9091SRichard Lowe * 4*5d9d9091SRichard Lowe * The contents of this file are subject to the terms of the 5*5d9d9091SRichard Lowe * Common Development and Distribution License (the "License"). 6*5d9d9091SRichard Lowe * You may not use this file except in compliance with the License. 7*5d9d9091SRichard Lowe * 8*5d9d9091SRichard Lowe * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9*5d9d9091SRichard Lowe * or http://www.opensolaris.org/os/licensing. 10*5d9d9091SRichard Lowe * See the License for the specific language governing permissions 11*5d9d9091SRichard Lowe * and limitations under the License. 12*5d9d9091SRichard Lowe * 13*5d9d9091SRichard Lowe * When distributing Covered Code, include this CDDL HEADER in each 14*5d9d9091SRichard Lowe * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15*5d9d9091SRichard Lowe * If applicable, add the following below this CDDL HEADER, with the 16*5d9d9091SRichard Lowe * fields enclosed by brackets "[]" replaced with your own identifying 17*5d9d9091SRichard Lowe * information: Portions Copyright [yyyy] [name of copyright owner] 18*5d9d9091SRichard Lowe * 19*5d9d9091SRichard Lowe * CDDL HEADER END 20*5d9d9091SRichard Lowe */ 21*5d9d9091SRichard Lowe 22*5d9d9091SRichard Lowe/* 23*5d9d9091SRichard Lowe * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 24*5d9d9091SRichard Lowe * Use is subject to license terms. 25*5d9d9091SRichard Lowe */ 26*5d9d9091SRichard Lowe 27*5d9d9091SRichard Lowe .file "strncmp.s" 28*5d9d9091SRichard Lowe 29*5d9d9091SRichard Lowe/* 30*5d9d9091SRichard Lowe * strncmp(s1, s2, n) 31*5d9d9091SRichard Lowe * 32*5d9d9091SRichard Lowe * Compare strings (at most n bytes): s1>s2: >0 s1==s2: 0 s1<s2: <0 33*5d9d9091SRichard Lowe * 34*5d9d9091SRichard Lowe * Fast assembler language version of the following C-program for strncmp 35*5d9d9091SRichard Lowe * which represents the `standard' for the C-library. 36*5d9d9091SRichard Lowe * 37*5d9d9091SRichard Lowe * int 38*5d9d9091SRichard Lowe * strncmp(const char *s1, const char *s2, size_t n) 39*5d9d9091SRichard Lowe * { 40*5d9d9091SRichard Lowe * n++; 41*5d9d9091SRichard Lowe * if (s1 == s2) 42*5d9d9091SRichard Lowe * return (0); 43*5d9d9091SRichard Lowe * while (--n != 0 && *s1 == *s2++) 44*5d9d9091SRichard Lowe * if (*s1++ == '\0') 45*5d9d9091SRichard Lowe * return (0); 46*5d9d9091SRichard Lowe * return ((n == 0) ? 0 : (*s1 - s2[-1])); 47*5d9d9091SRichard Lowe * } 48*5d9d9091SRichard Lowe */ 49*5d9d9091SRichard Lowe 50*5d9d9091SRichard Lowe#include <sys/asm_linkage.h> 51*5d9d9091SRichard Lowe 52*5d9d9091SRichard Lowe ENTRY(strncmp) 53*5d9d9091SRichard Lowe save %sp, -SA(WINDOWSIZE), %sp 54*5d9d9091SRichard Lowe cmp %i2, 8 55*5d9d9091SRichard Lowe blu,a .cmp_bytes ! for small counts go do bytes 56*5d9d9091SRichard Lowe sub %i0, %i1, %i0 ! delay slot, get diff from s1 - s2 57*5d9d9091SRichard Lowe andcc %i0, 3, %g0 ! is s1 aligned 58*5d9d9091SRichard Lowe1: bz .iss2 ! if so go check s2 59*5d9d9091SRichard Lowe andcc %i1, 3, %i3 ! is s2 aligned 60*5d9d9091SRichard Lowe 61*5d9d9091SRichard Lowe deccc %i2 ! --n >= 0 ? 62*5d9d9091SRichard Lowe bcs .doneq 63*5d9d9091SRichard Lowe nop ! delay slot 64*5d9d9091SRichard Lowe 65*5d9d9091SRichard Lowe ldub [%i0], %i4 ! else cmp one byte 66*5d9d9091SRichard Lowe ldub [%i1], %i5 67*5d9d9091SRichard Lowe inc %i0 68*5d9d9091SRichard Lowe cmp %i4, %i5 69*5d9d9091SRichard Lowe bne .noteqb 70*5d9d9091SRichard Lowe inc %i1 71*5d9d9091SRichard Lowe tst %i4 ! terminating zero 72*5d9d9091SRichard Lowe bnz 1b 73*5d9d9091SRichard Lowe andcc %i0, 3, %g0 74*5d9d9091SRichard Lowe b,a .doneq 75*5d9d9091SRichard Lowe 76*5d9d9091SRichard Lowe.iss2: 77*5d9d9091SRichard Lowe set 0x7efefeff, %l6 78*5d9d9091SRichard Lowe set 0x81010100, %l7 79*5d9d9091SRichard Lowe sethi %hi(0xff000000), %l0 ! masks to test for terminating null 80*5d9d9091SRichard Lowe sethi %hi(0x00ff0000), %l1 81*5d9d9091SRichard Lowe srl %l1, 8, %l2 ! generate 0x0000ff00 mask 82*5d9d9091SRichard Lowe 83*5d9d9091SRichard Lowe bz .w4cmp ! if s2 word aligned, compare words 84*5d9d9091SRichard Lowe cmp %i3, 2 ! check if s2 half aligned 85*5d9d9091SRichard Lowe be .w2cmp 86*5d9d9091SRichard Lowe cmp %i3, 1 ! check if aligned to 1 or 3 bytes 87*5d9d9091SRichard Lowe.w3cmp: ldub [%i1], %i5 88*5d9d9091SRichard Lowe inc 1, %i1 89*5d9d9091SRichard Lowe be .w1cmp 90*5d9d9091SRichard Lowe sll %i5, 24, %i5 91*5d9d9091SRichard Lowe sub %i0, %i1, %i0 92*5d9d9091SRichard Lowe2: 93*5d9d9091SRichard Lowe deccc 4, %i2 ! n >= 4 ? 94*5d9d9091SRichard Lowe bgeu,a 3f 95*5d9d9091SRichard Lowe ld [%i1], %i3 ! delay slot 96*5d9d9091SRichard Lowe dec %i1 ! reset s2 97*5d9d9091SRichard Lowe inc %i0 ! reset s1 diff 98*5d9d9091SRichard Lowe b .cmp_bytes ! do a byte at a time if n < 4 99*5d9d9091SRichard Lowe inc 4, %i2 100*5d9d9091SRichard Lowe3: 101*5d9d9091SRichard Lowe ld [%i0 + %i1], %i4 102*5d9d9091SRichard Lowe inc 4, %i1 103*5d9d9091SRichard Lowe srl %i3, 8, %l4 ! merge with the other half 104*5d9d9091SRichard Lowe or %l4, %i5, %i5 105*5d9d9091SRichard Lowe cmp %i4, %i5 106*5d9d9091SRichard Lowe be 1f 107*5d9d9091SRichard Lowe 108*5d9d9091SRichard Lowe add %i4, %l6, %l3 109*5d9d9091SRichard Lowe b,a .noteq 110*5d9d9091SRichard Lowe1: xor %l3, %i4, %l3 111*5d9d9091SRichard Lowe and %l3, %l7, %l3 112*5d9d9091SRichard Lowe cmp %l3, %l7 113*5d9d9091SRichard Lowe be,a 2b 114*5d9d9091SRichard Lowe sll %i3, 24, %i5 115*5d9d9091SRichard Lowe 116*5d9d9091SRichard Lowe ! 117*5d9d9091SRichard Lowe ! For 7-bit characters, we know one of the bytes is zero, but for 118*5d9d9091SRichard Lowe ! 8-bit characters, the zero detection algorithm gives some false 119*5d9d9091SRichard Lowe ! triggers ... check every byte individually. 120*5d9d9091SRichard Lowe ! 121*5d9d9091SRichard Lowe andcc %i4, %l0, %g0 ! check if first byte was zero 122*5d9d9091SRichard Lowe bnz 1f 123*5d9d9091SRichard Lowe andcc %i4, %l1, %g0 ! check if second byte was zero 124*5d9d9091SRichard Lowe b,a .doneq 125*5d9d9091SRichard Lowe1: bnz 1f 126*5d9d9091SRichard Lowe andcc %i4, %l2, %g0 ! check if third byte was zero 127*5d9d9091SRichard Lowe b,a .doneq 128*5d9d9091SRichard Lowe1: bnz 1f 129*5d9d9091SRichard Lowe andcc %i4, 0xff, %g0 ! check if last byte is zero 130*5d9d9091SRichard Lowe b,a .doneq 131*5d9d9091SRichard Lowe1: bnz 2b 132*5d9d9091SRichard Lowe sll %i3, 24, %i5 133*5d9d9091SRichard Lowe b,a .doneq 134*5d9d9091SRichard Lowe 135*5d9d9091SRichard Lowe.w1cmp: clr %l4 136*5d9d9091SRichard Lowe lduh [%i1], %l4 137*5d9d9091SRichard Lowe inc 2, %i1 138*5d9d9091SRichard Lowe sll %l4, 8, %l4 139*5d9d9091SRichard Lowe or %i5, %l4, %i5 140*5d9d9091SRichard Lowe 141*5d9d9091SRichard Lowe sub %i0, %i1, %i0 142*5d9d9091SRichard Lowe3: 143*5d9d9091SRichard Lowe deccc 4, %i2 ! n >= 4 ? 144*5d9d9091SRichard Lowe bgeu,a 4f 145*5d9d9091SRichard Lowe ld [%i1], %i3 ! delay slot 146*5d9d9091SRichard Lowe dec 3, %i1 ! reset s2 147*5d9d9091SRichard Lowe inc 3, %i0 ! reset s1 diff 148*5d9d9091SRichard Lowe b .cmp_bytes ! do a byte at a time if n < 4 149*5d9d9091SRichard Lowe inc 4, %i2 150*5d9d9091SRichard Lowe4: 151*5d9d9091SRichard Lowe ld [%i0 + %i1], %i4 152*5d9d9091SRichard Lowe inc 4, %i1 153*5d9d9091SRichard Lowe srl %i3, 24, %l4 ! merge with the other half 154*5d9d9091SRichard Lowe or %l4, %i5, %i5 155*5d9d9091SRichard Lowe cmp %i4, %i5 156*5d9d9091SRichard Lowe be 1f 157*5d9d9091SRichard Lowe 158*5d9d9091SRichard Lowe add %i4, %l6, %l3 159*5d9d9091SRichard Lowe b,a .noteq 160*5d9d9091SRichard Lowe1: xor %l3, %i4, %l3 161*5d9d9091SRichard Lowe and %l3, %l7, %l3 162*5d9d9091SRichard Lowe cmp %l3, %l7 163*5d9d9091SRichard Lowe be,a 3b 164*5d9d9091SRichard Lowe sll %i3, 8, %i5 165*5d9d9091SRichard Lowe 166*5d9d9091SRichard Lowe andcc %i4, %l0, %g0 ! check if first byte was zero 167*5d9d9091SRichard Lowe bnz 1f 168*5d9d9091SRichard Lowe andcc %i4, %l1, %g0 ! check if second byte was zero 169*5d9d9091SRichard Lowe b,a .doneq 170*5d9d9091SRichard Lowe1: bnz 1f 171*5d9d9091SRichard Lowe andcc %i4, %l2, %g0 ! check if third byte was zero 172*5d9d9091SRichard Lowe b,a .doneq 173*5d9d9091SRichard Lowe1: bnz 1f 174*5d9d9091SRichard Lowe andcc %i4, 0xff, %g0 ! check if last byte is zero 175*5d9d9091SRichard Lowe b,a .doneq 176*5d9d9091SRichard Lowe1: bnz 3b 177*5d9d9091SRichard Lowe sll %i3, 8, %i5 178*5d9d9091SRichard Lowe b,a .doneq 179*5d9d9091SRichard Lowe 180*5d9d9091SRichard Lowe.w2cmp: 181*5d9d9091SRichard Lowe lduh [%i1], %i5 ! read a halfword to align s2 182*5d9d9091SRichard Lowe inc 2, %i1 183*5d9d9091SRichard Lowe sll %i5, 16, %i5 184*5d9d9091SRichard Lowe 185*5d9d9091SRichard Lowe sub %i0, %i1, %i0 186*5d9d9091SRichard Lowe4: 187*5d9d9091SRichard Lowe deccc 4, %i2 ! n >= 4 ? 188*5d9d9091SRichard Lowe bgeu,a 5f 189*5d9d9091SRichard Lowe ld [%i1], %i3 ! delay slot 190*5d9d9091SRichard Lowe dec 2, %i1 ! reset s2 191*5d9d9091SRichard Lowe inc 2, %i0 ! reset s1 diff 192*5d9d9091SRichard Lowe b .cmp_bytes ! do a byte at a time if n < 4 193*5d9d9091SRichard Lowe inc 4, %i2 ! delay slot 194*5d9d9091SRichard Lowe5: 195*5d9d9091SRichard Lowe ld [%i1 + %i0], %i4 ! read a word from s2 196*5d9d9091SRichard Lowe inc 4, %i1 197*5d9d9091SRichard Lowe srl %i3, 16, %l4 ! merge with the other half 198*5d9d9091SRichard Lowe or %l4, %i5, %i5 199*5d9d9091SRichard Lowe cmp %i4, %i5 200*5d9d9091SRichard Lowe be 1f 201*5d9d9091SRichard Lowe 202*5d9d9091SRichard Lowe add %i4, %l6, %l3 203*5d9d9091SRichard Lowe b,a .noteq 204*5d9d9091SRichard Lowe1: xor %l3, %i4, %l3 ! are any bytes 0? 205*5d9d9091SRichard Lowe and %l3, %l7, %l3 206*5d9d9091SRichard Lowe cmp %l3, %l7 207*5d9d9091SRichard Lowe be,a 4b 208*5d9d9091SRichard Lowe sll %i3, 16, %i5 209*5d9d9091SRichard Lowe 210*5d9d9091SRichard Lowe andcc %i4, %l0, %g0 ! check if first byte was zero 211*5d9d9091SRichard Lowe bnz 1f 212*5d9d9091SRichard Lowe andcc %i4, %l1, %g0 ! check if second byte was zero 213*5d9d9091SRichard Lowe b,a .doneq 214*5d9d9091SRichard Lowe1: bnz 1f 215*5d9d9091SRichard Lowe andcc %i4, %l2, %g0 ! check if third byte was zero 216*5d9d9091SRichard Lowe b,a .doneq 217*5d9d9091SRichard Lowe1: bnz 1f 218*5d9d9091SRichard Lowe andcc %i4, 0xff, %g0 ! check if last byte is zero 219*5d9d9091SRichard Lowe b,a .doneq 220*5d9d9091SRichard Lowe1: bnz 4b 221*5d9d9091SRichard Lowe sll %i3, 16, %i5 222*5d9d9091SRichard Lowe b,a .doneq 223*5d9d9091SRichard Lowe 224*5d9d9091SRichard Lowe.w4cmp: sub %i0, %i1, %i0 225*5d9d9091SRichard Lowe ld [%i1], %i5 ! read a word from s1 226*5d9d9091SRichard Lowe5: cmp %i2,0 227*5d9d9091SRichard Lowe be,a .doneq 228*5d9d9091SRichard Lowe nop 229*5d9d9091SRichard Lowe ld [%i1], %i5 ! read a word from s1 230*5d9d9091SRichard Lowe deccc 4, %i2 ! n >= 4 ? 231*5d9d9091SRichard Lowe bcs,a .cmp_bytes ! do a byte at a time if n < 4 232*5d9d9091SRichard Lowe inc 4, %i2 233*5d9d9091SRichard Lowe 234*5d9d9091SRichard Lowe ld [%i1 + %i0], %i4 ! read a word from s2 235*5d9d9091SRichard Lowe cmp %i4, %i5 236*5d9d9091SRichard Lowe inc 4, %i1 237*5d9d9091SRichard Lowe be 1f 238*5d9d9091SRichard Lowe 239*5d9d9091SRichard Lowe add %i4, %l6, %l3 240*5d9d9091SRichard Lowe b,a .noteq 241*5d9d9091SRichard Lowe1: xor %l3, %i4, %l3 242*5d9d9091SRichard Lowe and %l3, %l7, %l3 243*5d9d9091SRichard Lowe cmp %l3, %l7 244*5d9d9091SRichard Lowe be,a 5b 245*5d9d9091SRichard Lowe nop 246*5d9d9091SRichard Lowe 247*5d9d9091SRichard Lowe andcc %i4, %l0, %g0 ! check if first byte was zero 248*5d9d9091SRichard Lowe bnz 1f 249*5d9d9091SRichard Lowe andcc %i4, %l1, %g0 ! check if second byte was zero 250*5d9d9091SRichard Lowe b,a .doneq 251*5d9d9091SRichard Lowe1: bnz 1f 252*5d9d9091SRichard Lowe andcc %i4, %l2, %g0 ! check if third byte was zero 253*5d9d9091SRichard Lowe b,a .doneq 254*5d9d9091SRichard Lowe1: bnz 1f 255*5d9d9091SRichard Lowe andcc %i4, 0xff, %g0 ! check if last byte is zero 256*5d9d9091SRichard Lowe b,a .doneq 257*5d9d9091SRichard Lowe1: bnz,a 5b 258*5d9d9091SRichard Lowe ld [%i1], %i5 259*5d9d9091SRichard Lowe.doneq: ret 260*5d9d9091SRichard Lowe restore %g0, %g0, %o0 ! equal return zero 261*5d9d9091SRichard Lowe 262*5d9d9091SRichard Lowe.noteq: srl %i4, 24, %l4 263*5d9d9091SRichard Lowe srl %i5, 24, %l5 264*5d9d9091SRichard Lowe subcc %l4, %l5, %i0 265*5d9d9091SRichard Lowe bne 6f 266*5d9d9091SRichard Lowe andcc %l4, 0xff, %g0 267*5d9d9091SRichard Lowe bz .doneq 268*5d9d9091SRichard Lowe sll %i4, 8, %l4 269*5d9d9091SRichard Lowe sll %i5, 8, %l5 270*5d9d9091SRichard Lowe srl %l4, 24, %l4 271*5d9d9091SRichard Lowe srl %l5, 24, %l5 272*5d9d9091SRichard Lowe subcc %l4, %l5, %i0 273*5d9d9091SRichard Lowe bne 6f 274*5d9d9091SRichard Lowe andcc %l4, 0xff, %g0 275*5d9d9091SRichard Lowe bz .doneq 276*5d9d9091SRichard Lowe sll %i4, 16, %l4 277*5d9d9091SRichard Lowe sll %i5, 16, %l5 278*5d9d9091SRichard Lowe srl %l4, 24, %l4 279*5d9d9091SRichard Lowe srl %l5, 24, %l5 280*5d9d9091SRichard Lowe subcc %l4, %l5, %i0 281*5d9d9091SRichard Lowe bne 6f 282*5d9d9091SRichard Lowe andcc %l4, 0xff, %g0 283*5d9d9091SRichard Lowe bz .doneq 284*5d9d9091SRichard Lowe nop 285*5d9d9091SRichard Lowe.noteqb: 286*5d9d9091SRichard Lowe and %i4, 0xff, %l4 287*5d9d9091SRichard Lowe and %i5, 0xff, %l5 288*5d9d9091SRichard Lowe subcc %l4, %l5, %i0 289*5d9d9091SRichard Lowe6: ret 290*5d9d9091SRichard Lowe restore %i0, %g0, %o0 291*5d9d9091SRichard Lowe 292*5d9d9091SRichard Lowe ! Do a byte by byte comparison, disregarding alignments 293*5d9d9091SRichard Lowe.cmp_bytes: 294*5d9d9091SRichard Lowe deccc %i2 ! --n >= 0 ? 295*5d9d9091SRichard Lowe1: 296*5d9d9091SRichard Lowe bcs .doneq 297*5d9d9091SRichard Lowe nop ! delay slot 298*5d9d9091SRichard Lowe ldub [%i1 + %i0], %i4 ! read a word from s1 299*5d9d9091SRichard Lowe ldub [%i1], %i5 ! read a word from s2 300*5d9d9091SRichard Lowe 301*5d9d9091SRichard Lowe inc %i1 302*5d9d9091SRichard Lowe cmp %i4, %i5 303*5d9d9091SRichard Lowe bne .noteqb 304*5d9d9091SRichard Lowe tst %i4 ! terminating zero 305*5d9d9091SRichard Lowe bnz 1b 306*5d9d9091SRichard Lowe deccc %i2 ! --n >= 0 307*5d9d9091SRichard Lowe b,a .doneq 308*5d9d9091SRichard Lowe 309*5d9d9091SRichard Lowe SET_SIZE(strncmp) 310