10Sstevel@tonic-gate/* 20Sstevel@tonic-gate * CDDL HEADER START 30Sstevel@tonic-gate * 40Sstevel@tonic-gate * The contents of this file are subject to the terms of the 56812Sraf * Common Development and Distribution License (the "License"). 66812Sraf * You may not use this file except in compliance with the License. 70Sstevel@tonic-gate * 80Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 90Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing. 100Sstevel@tonic-gate * See the License for the specific language governing permissions 110Sstevel@tonic-gate * and limitations under the License. 120Sstevel@tonic-gate * 130Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each 140Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 150Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the 160Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying 170Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner] 180Sstevel@tonic-gate * 190Sstevel@tonic-gate * CDDL HEADER END 200Sstevel@tonic-gate */ 216812Sraf 220Sstevel@tonic-gate/* 236812Sraf * Copyright 2008 Sun Microsystems, Inc. All rights reserved. 240Sstevel@tonic-gate * Use is subject to license terms. 250Sstevel@tonic-gate */ 260Sstevel@tonic-gate 27*7298SMark.J.Nelson@Sun.COM .file "strcpy.s" 280Sstevel@tonic-gate 290Sstevel@tonic-gate/* 300Sstevel@tonic-gate * strcpy(s1, s2) 310Sstevel@tonic-gate * 320Sstevel@tonic-gate * Copy string s2 to s1. s1 must be large enough. Return s1. 330Sstevel@tonic-gate * 340Sstevel@tonic-gate * Fast assembler language version of the following C-program strcpy 350Sstevel@tonic-gate * which represents the `standard' for the C-library. 360Sstevel@tonic-gate * 370Sstevel@tonic-gate * char * 380Sstevel@tonic-gate * strcpy(s1, s2) 390Sstevel@tonic-gate * register char *s1; 400Sstevel@tonic-gate * register const char *s2; 410Sstevel@tonic-gate * { 420Sstevel@tonic-gate * char *os1 = s1; 430Sstevel@tonic-gate * 440Sstevel@tonic-gate * while(*s1++ = *s2++) 450Sstevel@tonic-gate * ; 460Sstevel@tonic-gate * return(os1); 470Sstevel@tonic-gate * } 480Sstevel@tonic-gate * 490Sstevel@tonic-gate */ 500Sstevel@tonic-gate 510Sstevel@tonic-gate#include <sys/asm_linkage.h> 520Sstevel@tonic-gate 530Sstevel@tonic-gate ! This is a 32-bit implementation of strcpy. It works by 540Sstevel@tonic-gate ! first checking the alignment of its source pointer. And, 550Sstevel@tonic-gate ! if it is not aligned, attempts to copy bytes until it is. 560Sstevel@tonic-gate ! once this has occurred, the copy takes place, while checking 570Sstevel@tonic-gate ! for zero bytes, based upon destination alignment. 580Sstevel@tonic-gate ! Methods exist to handle per-byte, half-word, and word sized 590Sstevel@tonic-gate ! copies. 600Sstevel@tonic-gate 610Sstevel@tonic-gate ENTRY(strcpy) 620Sstevel@tonic-gate 630Sstevel@tonic-gate .align 32 640Sstevel@tonic-gate 650Sstevel@tonic-gate sub %o1, %o0, %o3 ! src - dst 660Sstevel@tonic-gate andcc %o1, 3, %o4 ! src word aligned ? 670Sstevel@tonic-gate bz .srcaligned ! yup 680Sstevel@tonic-gate mov %o0, %o2 ! save dst 690Sstevel@tonic-gate 700Sstevel@tonic-gate cmp %o4, 2 ! src halfword aligned 710Sstevel@tonic-gate be .s2aligned ! yup 720Sstevel@tonic-gate ldub [%o2 + %o3], %o1 ! src[0] 730Sstevel@tonic-gate tst %o1 ! byte zero? 740Sstevel@tonic-gate stb %o1, [%o2] ! store first byte 750Sstevel@tonic-gate bz .done ! yup, done 760Sstevel@tonic-gate cmp %o4, 3 ! only one byte needed to align? 770Sstevel@tonic-gate bz .srcaligned ! yup 780Sstevel@tonic-gate inc %o2 ! src++, dst++ 790Sstevel@tonic-gate 800Sstevel@tonic-gate.s2aligned: 810Sstevel@tonic-gate lduh [%o2 + %o3], %o1 ! src[] 820Sstevel@tonic-gate srl %o1, 8, %o4 ! %o4<7:0> = first byte 830Sstevel@tonic-gate tst %o4 ! first byte zero ? 840Sstevel@tonic-gate bz .done ! yup, done 850Sstevel@tonic-gate stb %o4, [%o2] ! store first byte 860Sstevel@tonic-gate andcc %o1, 0xff, %g0 ! second byte zero ? 870Sstevel@tonic-gate bz .done ! yup, done 880Sstevel@tonic-gate stb %o1, [%o2 + 1] ! store second byte 890Sstevel@tonic-gate add %o2, 2, %o2 ! src += 2, dst += 2 900Sstevel@tonic-gate 910Sstevel@tonic-gate.srcaligned: 920Sstevel@tonic-gate sethi %hi(0x01010101), %o4 ! Alan Mycroft's magic1 930Sstevel@tonic-gate sethi %hi(0x80808080), %o5 ! Alan Mycroft's magic2 940Sstevel@tonic-gate or %o4, %lo(0x01010101), %o4 950Sstevel@tonic-gate andcc %o2, 3, %o1 ! destination word aligned? 960Sstevel@tonic-gate bnz .dstnotaligned ! nope 970Sstevel@tonic-gate or %o5, %lo(0x80808080), %o5 980Sstevel@tonic-gate 990Sstevel@tonic-gate.copyword: 1000Sstevel@tonic-gate lduw [%o2 + %o3], %o1 ! src word 1010Sstevel@tonic-gate add %o2, 4, %o2 ! src += 4, dst += 4 1020Sstevel@tonic-gate andn %o5, %o1, %g1 ! ~word & 0x80808080 1030Sstevel@tonic-gate sub %o1, %o4, %o1 ! word - 0x01010101 1040Sstevel@tonic-gate andcc %o1, %g1, %g0 ! ((word - 0x01010101) & ~word & 0x80808080) 1050Sstevel@tonic-gate add %o1, %o4, %o1 ! restore word 1060Sstevel@tonic-gate bz,a .copyword ! no zero byte if magic expression == 0 1070Sstevel@tonic-gate st %o1, [%o2 - 4] ! store word to dst (address pre-incremented) 1080Sstevel@tonic-gate 1090Sstevel@tonic-gate.zerobyte: 1100Sstevel@tonic-gate set 0xff000000, %o4 ! mask for 1st byte 1110Sstevel@tonic-gate srl %o1, 24, %o3 ! %o3<7:0> = first byte 1120Sstevel@tonic-gate andcc %o1, %o4, %g0 ! first byte zero? 1130Sstevel@tonic-gate bz .done ! yup, done 1140Sstevel@tonic-gate stb %o3, [%o2 - 4] ! store first byte 1150Sstevel@tonic-gate set 0x00ff0000, %o5 ! mask for 2nd byte 1160Sstevel@tonic-gate srl %o1, 16, %o3 ! %o3<7:0> = second byte 1170Sstevel@tonic-gate andcc %o1, %o5, %g0 ! second byte zero? 1180Sstevel@tonic-gate bz .done ! yup, done 1190Sstevel@tonic-gate stb %o3, [%o2 - 3] ! store second byte 1200Sstevel@tonic-gate srl %o4, 16, %o4 ! 0x0000ff00 = mask for 3rd byte 1210Sstevel@tonic-gate andcc %o1, %o4, %g0 ! third byte zero? 1220Sstevel@tonic-gate srl %o1, 8, %o3 ! %o3<7:0> = third byte 1230Sstevel@tonic-gate bz .done ! yup, done 1240Sstevel@tonic-gate stb %o3, [%o2 - 2] ! store third byte 1250Sstevel@tonic-gate stb %o1, [%o2 - 1] ! store fourth byte 1260Sstevel@tonic-gate 1270Sstevel@tonic-gate.done: 1280Sstevel@tonic-gate retl ! done with leaf function 1290Sstevel@tonic-gate .empty 1300Sstevel@tonic-gate 1310Sstevel@tonic-gate.dstnotaligned: 1320Sstevel@tonic-gate cmp %o1, 2 ! dst half word aligned? 1330Sstevel@tonic-gate be,a .storehalfword2 ! yup, store half word at a time 1340Sstevel@tonic-gate lduw [%o2 + %o3], %o1 ! src word 1350Sstevel@tonic-gate 1360Sstevel@tonic-gate.storebyte: 1370Sstevel@tonic-gate lduw [%o2 + %o3], %o1 ! src word 1380Sstevel@tonic-gate add %o2, 4, %o2 ! src += 4, dst += 4 1390Sstevel@tonic-gate sub %o1, %o4, %g1 ! x - 0x01010101 1400Sstevel@tonic-gate andn %g1, %o1, %g1 ! (x - 0x01010101) & ~x 1410Sstevel@tonic-gate andcc %g1, %o5, %g0 ! ((x - 0x01010101) & ~x & 0x80808080) 1420Sstevel@tonic-gate bnz .zerobyte ! word has zero byte, handle end cases 1430Sstevel@tonic-gate srl %o1, 24, %g1 ! %g1<7:0> = first byte 1440Sstevel@tonic-gate stb %g1, [%o2 - 4] ! store first byte; half-word aligned now 1450Sstevel@tonic-gate srl %o1, 8, %g1 ! %g1<15:0> = byte 2, 3 1460Sstevel@tonic-gate sth %g1, [%o2 - 3] ! store bytes 2, 3 1470Sstevel@tonic-gate ba .storebyte ! next word 1480Sstevel@tonic-gate stb %o1, [%o2 - 1] ! store fourth byte 1490Sstevel@tonic-gate 1500Sstevel@tonic-gate.storehalfword: 1510Sstevel@tonic-gate lduw [%o2 + %o3], %o1 ! src word 1520Sstevel@tonic-gate.storehalfword2: 1530Sstevel@tonic-gate add %o2, 4, %o2 ! src += 4, dst += 4 1540Sstevel@tonic-gate sub %o1, %o4, %g1 ! x - 0x01010101 1550Sstevel@tonic-gate andn %g1, %o1, %g1 ! (x - 0x01010101) & ~x 1560Sstevel@tonic-gate andcc %g1, %o5, %g0 ! ((x - 0x01010101) & ~x & 0x80808080) 1570Sstevel@tonic-gate bnz .zerobyte ! word has zero byte, handle end cases 1580Sstevel@tonic-gate srl %o1, 16, %g1 ! get first and second byte 1590Sstevel@tonic-gate sth %g1, [%o2 - 4] ! store first and second byte 1600Sstevel@tonic-gate ba .storehalfword ! next word 1610Sstevel@tonic-gate sth %o1, [%o2 - 2] ! store third and fourth byte 1620Sstevel@tonic-gate 1630Sstevel@tonic-gate ! DO NOT remove these NOPs. It will slow down the halfword loop by 15% 1640Sstevel@tonic-gate 1650Sstevel@tonic-gate nop ! padding 1660Sstevel@tonic-gate nop ! padding 1670Sstevel@tonic-gate 1680Sstevel@tonic-gate SET_SIZE(strcpy) 1690Sstevel@tonic-gate 170