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 2009 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 "memcpy.s" 28*5d9d9091SRichard Lowe 29*5d9d9091SRichard Lowe#include <sys/asm_linkage.h> 30*5d9d9091SRichard Lowe 31*5d9d9091SRichard Lowe ANSI_PRAGMA_WEAK(memmove,function) 32*5d9d9091SRichard Lowe ANSI_PRAGMA_WEAK(memcpy,function) 33*5d9d9091SRichard Lowe 34*5d9d9091SRichard Lowe#include "SYS.h" 35*5d9d9091SRichard Lowe 36*5d9d9091SRichard Lowe ENTRY(memcpy) 37*5d9d9091SRichard Lowe movl %edi,%edx / save register variables 38*5d9d9091SRichard Lowe pushl %esi 39*5d9d9091SRichard Lowe movl 8(%esp),%edi / %edi = dest address 40*5d9d9091SRichard Lowe movl 12(%esp),%esi / %esi = source address 41*5d9d9091SRichard Lowe movl 16(%esp),%ecx / %ecx = length of string 42*5d9d9091SRichard Lowe movl %edi,%eax / return value from the call 43*5d9d9091SRichard Lowe 44*5d9d9091SRichard Lowe shrl $2,%ecx / %ecx = number of words to move 45*5d9d9091SRichard Lowe rep ; smovl / move the words 46*5d9d9091SRichard Lowe 47*5d9d9091SRichard Lowe movl 16(%esp),%ecx / %ecx = number of bytes to move 48*5d9d9091SRichard Lowe andl $0x3,%ecx / %ecx = number of bytes left to move 49*5d9d9091SRichard Lowe rep ; smovb / move the bytes 50*5d9d9091SRichard Lowe 51*5d9d9091SRichard Lowe popl %esi / restore register variables 52*5d9d9091SRichard Lowe movl %edx,%edi 53*5d9d9091SRichard Lowe ret 54*5d9d9091SRichard Lowe SET_SIZE(memcpy) 55*5d9d9091SRichard Lowe 56*5d9d9091SRichard Lowe 57*5d9d9091SRichard Lowe ENTRY(memmove) 58*5d9d9091SRichard Lowe pushl %edi / save off %edi, %esi and move destination 59*5d9d9091SRichard Lowe movl 4+12(%esp),%ecx / get number of bytes to move 60*5d9d9091SRichard Lowe pushl %esi 61*5d9d9091SRichard Lowe testl %ecx,%ecx / if (n == 0) 62*5d9d9091SRichard Lowe je .CleanupReturn / return(s); 63*5d9d9091SRichard Lowe movl 8+ 4(%esp),%edi / destination buffer address 64*5d9d9091SRichard Lowe movl 8+ 8(%esp),%esi / source buffer address 65*5d9d9091SRichard Lowe.Common: 66*5d9d9091SRichard Lowe movl $3,%eax / heavily used constant 67*5d9d9091SRichard Lowe cmpl %esi,%edi / if (source addr > dest addr) 68*5d9d9091SRichard Lowe leal -1(%esi,%ecx),%edx 69*5d9d9091SRichard Lowe jbe .CopyRight / 70*5d9d9091SRichard Lowe cmpl %edx,%edi 71*5d9d9091SRichard Lowe jbe .CopyLeft 72*5d9d9091SRichard Lowe.CopyRight: 73*5d9d9091SRichard Lowe cmpl $8,%ecx / if (size < 8 bytes) 74*5d9d9091SRichard Lowe jbe .OneByteCopy / goto fast short copy loop 75*5d9d9091SRichard Lowe.FourByteCopy: 76*5d9d9091SRichard Lowe movl %ecx,%edx / save count 77*5d9d9091SRichard Lowe movl %esi,%ecx / get source buffer 4 byte aligned 78*5d9d9091SRichard Lowe andl %eax,%ecx 79*5d9d9091SRichard Lowe jz .SkipAlignRight 80*5d9d9091SRichard Lowe subl %ecx,%edx 81*5d9d9091SRichard Lowe rep; smovb / do the byte part of copy 82*5d9d9091SRichard Lowe.SkipAlignRight: 83*5d9d9091SRichard Lowe movl %edx,%ecx 84*5d9d9091SRichard Lowe shrl $2,%ecx 85*5d9d9091SRichard Lowe rep; smovl / do the long word part 86*5d9d9091SRichard Lowe movl %edx,%ecx / compute bytes left to move 87*5d9d9091SRichard Lowe andl %eax,%ecx / complete copy of remaining bytes 88*5d9d9091SRichard Lowe jz .CleanupReturn 89*5d9d9091SRichard Lowe.OneByteCopy: 90*5d9d9091SRichard Lowe rep; smovb / do the byte part of copy 91*5d9d9091SRichard Lowe.CleanupReturn: 92*5d9d9091SRichard Lowe popl %esi / } 93*5d9d9091SRichard Lowe popl %edi / restore registers 94*5d9d9091SRichard Lowe movl 4(%esp),%eax / set up return value 95*5d9d9091SRichard Lowe.Return: 96*5d9d9091SRichard Lowe ret / return(dba); 97*5d9d9091SRichard Lowe 98*5d9d9091SRichard Lowe.CopyLeft: 99*5d9d9091SRichard Lowe std / reverse direction bit (RtoL) 100*5d9d9091SRichard Lowe cmpl $12,%ecx / if (size < 12) 101*5d9d9091SRichard Lowe ja .BigCopyLeft / { 102*5d9d9091SRichard Lowe movl %edx,%esi / src = src + size - 1 103*5d9d9091SRichard Lowe leal -1(%ecx,%edi),%edi / dst = dst + size - 1 104*5d9d9091SRichard Lowe rep; smovb / do the byte copy 105*5d9d9091SRichard Lowe cld / reset direction flag to LtoR 106*5d9d9091SRichard Lowe popl %esi / } 107*5d9d9091SRichard Lowe popl %edi / restore registers 108*5d9d9091SRichard Lowe movl 4(%esp),%eax / set up return value 109*5d9d9091SRichard Lowe ret / return(dba); 110*5d9d9091SRichard Lowe.BigCopyLeft: / } else { 111*5d9d9091SRichard Lowe xchgl %edx,%ecx 112*5d9d9091SRichard Lowe movl %ecx,%esi / align source w/byte copy 113*5d9d9091SRichard Lowe leal -1(%edx,%edi),%edi 114*5d9d9091SRichard Lowe andl %eax,%ecx 115*5d9d9091SRichard Lowe jz .SkipAlignLeft 116*5d9d9091SRichard Lowe addl $1, %ecx / we need to insure that future 117*5d9d9091SRichard Lowe subl %ecx,%edx / copy is done on aligned boundary 118*5d9d9091SRichard Lowe rep; smovb 119*5d9d9091SRichard Lowe.SkipAlignLeft: 120*5d9d9091SRichard Lowe movl %edx,%ecx 121*5d9d9091SRichard Lowe subl %eax,%esi 122*5d9d9091SRichard Lowe shrl $2,%ecx / do 4 byte copy RtoL 123*5d9d9091SRichard Lowe subl %eax,%edi 124*5d9d9091SRichard Lowe rep; smovl 125*5d9d9091SRichard Lowe andl %eax,%edx / do 1 byte copy whats left 126*5d9d9091SRichard Lowe jz .CleanupReturnLeft 127*5d9d9091SRichard Lowe movl %edx,%ecx 128*5d9d9091SRichard Lowe addl %eax,%esi / rep; smovl instruction will decrement 129*5d9d9091SRichard Lowe addl %eax,%edi / %edi, %esi by four after each copy 130*5d9d9091SRichard Lowe / adding 3 will restore pointers to byte 131*5d9d9091SRichard Lowe / before last double word copied 132*5d9d9091SRichard Lowe / which is where they are expected to 133*5d9d9091SRichard Lowe / be for the single byte copy code 134*5d9d9091SRichard Lowe rep; smovb 135*5d9d9091SRichard Lowe.CleanupReturnLeft: 136*5d9d9091SRichard Lowe cld / reset direction flag to LtoR 137*5d9d9091SRichard Lowe popl %esi 138*5d9d9091SRichard Lowe popl %edi / restore registers 139*5d9d9091SRichard Lowe movl 4(%esp),%eax / set up return value 140*5d9d9091SRichard Lowe ret / return(dba); 141*5d9d9091SRichard Lowe SET_SIZE(memmove) 142