startprog64.S revision 1.3.2.2
1 1.3.2.2 pgoyette /* $NetBSD: startprog64.S,v 1.3.2.2 2017/03/20 06:57:15 pgoyette Exp $ */ 2 1.3.2.2 pgoyette /* NetBSD: startprog.S,v 1.3 2003/02/01 14:48:18 dsl Exp */ 3 1.3.2.2 pgoyette 4 1.3.2.2 pgoyette /* starts program in protected mode / flat space 5 1.3.2.2 pgoyette with given stackframe 6 1.3.2.2 pgoyette needs global variables flatcodeseg and flatdataseg 7 1.3.2.2 pgoyette (gdt offsets) 8 1.3.2.2 pgoyette derived from: NetBSD:sys/arch/i386/boot/asm.S 9 1.3.2.2 pgoyette */ 10 1.3.2.2 pgoyette 11 1.3.2.2 pgoyette /* 12 1.3.2.2 pgoyette * Ported to boot 386BSD by Julian Elischer (julian (at) tfs.com) Sept 1992 13 1.3.2.2 pgoyette * 14 1.3.2.2 pgoyette * Mach Operating System 15 1.3.2.2 pgoyette * Copyright (c) 1992, 1991 Carnegie Mellon University 16 1.3.2.2 pgoyette * All Rights Reserved. 17 1.3.2.2 pgoyette * 18 1.3.2.2 pgoyette * Permission to use, copy, modify and distribute this software and its 19 1.3.2.2 pgoyette * documentation is hereby granted, provided that both the copyright 20 1.3.2.2 pgoyette * notice and this permission notice appear in all copies of the 21 1.3.2.2 pgoyette * software, derivative works or modified versions, and any portions 22 1.3.2.2 pgoyette * thereof, and that both notices appear in supporting documentation. 23 1.3.2.2 pgoyette * 24 1.3.2.2 pgoyette * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 25 1.3.2.2 pgoyette * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR 26 1.3.2.2 pgoyette * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 27 1.3.2.2 pgoyette * 28 1.3.2.2 pgoyette * Carnegie Mellon requests users of this software to return to 29 1.3.2.2 pgoyette * 30 1.3.2.2 pgoyette * Software Distribution Coordinator or Software.Distribution (at) CS.CMU.EDU 31 1.3.2.2 pgoyette * School of Computer Science 32 1.3.2.2 pgoyette * Carnegie Mellon University 33 1.3.2.2 pgoyette * Pittsburgh PA 15213-3890 34 1.3.2.2 pgoyette * 35 1.3.2.2 pgoyette * any improvements or extensions that they make and grant Carnegie Mellon 36 1.3.2.2 pgoyette * the rights to redistribute these changes. 37 1.3.2.2 pgoyette */ 38 1.3.2.2 pgoyette 39 1.3.2.2 pgoyette /* 40 1.3.2.2 pgoyette Copyright 1988, 1989, 1990, 1991, 1992 41 1.3.2.2 pgoyette by Intel Corporation, Santa Clara, California. 42 1.3.2.2 pgoyette 43 1.3.2.2 pgoyette All Rights Reserved 44 1.3.2.2 pgoyette 45 1.3.2.2 pgoyette Permission to use, copy, modify, and distribute this software and 46 1.3.2.2 pgoyette its documentation for any purpose and without fee is hereby 47 1.3.2.2 pgoyette granted, provided that the above copyright notice appears in all 48 1.3.2.2 pgoyette copies and that both the copyright notice and this permission notice 49 1.3.2.2 pgoyette appear in supporting documentation, and that the name of Intel 50 1.3.2.2 pgoyette not be used in advertising or publicity pertaining to distribution 51 1.3.2.2 pgoyette of the software without specific, written prior permission. 52 1.3.2.2 pgoyette 53 1.3.2.2 pgoyette INTEL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE 54 1.3.2.2 pgoyette INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, 55 1.3.2.2 pgoyette IN NO EVENT SHALL INTEL BE LIABLE FOR ANY SPECIAL, INDIRECT, OR 56 1.3.2.2 pgoyette CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM 57 1.3.2.2 pgoyette LOSS OF USE, DATA OR PROFITS, WHETHER IN ACTION OF CONTRACT, 58 1.3.2.2 pgoyette NEGLIGENCE, OR OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION 59 1.3.2.2 pgoyette WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. 60 1.3.2.2 pgoyette */ 61 1.3.2.2 pgoyette 62 1.3.2.2 pgoyette #include <machine/asm.h> 63 1.3.2.2 pgoyette #include <machine/specialreg.h> 64 1.3.2.2 pgoyette 65 1.3.2.2 pgoyette #define CODE_SEGMENT 0x08 66 1.3.2.2 pgoyette #define DATA_SEGMENT 0x10 67 1.3.2.2 pgoyette 68 1.3.2.2 pgoyette .align 16 69 1.3.2.2 pgoyette .globl _C_LABEL(startprog64) 70 1.3.2.2 pgoyette _C_LABEL(startprog64): 71 1.3.2.2 pgoyette .quad 0 72 1.3.2.2 pgoyette 73 1.3.2.2 pgoyette .globl _C_LABEL(startprog64_size) 74 1.3.2.2 pgoyette _C_LABEL(startprog64_size): 75 1.3.2.2 pgoyette .long startprog64_end - _C_LABEL(startprog64_start) 76 1.3.2.2 pgoyette 77 1.3.2.2 pgoyette .text 78 1.3.2.2 pgoyette .p2align 4,,15 79 1.3.2.2 pgoyette 80 1.3.2.2 pgoyette /* 81 1.3.2.2 pgoyette * startprog64(loadddr,entry,stack,kern_load,kern_start,kern_size) 82 1.3.2.2 pgoyette */ 83 1.3.2.2 pgoyette ENTRY(startprog64_start) 84 1.3.2.2 pgoyette start: 85 1.3.2.2 pgoyette /* 86 1.3.2.2 pgoyette * This function is to call the loaded kernel's start() with 87 1.3.2.2 pgoyette * 32bit segment mode from x64 mode. 88 1.3.2.2 pgoyette * %rdi: kernel start address 89 1.3.2.2 pgoyette * %rsi: loaded kernel address 90 1.3.2.2 pgoyette * %rdx: stack address 91 1.3.2.2 pgoyette * %rcx: loaded kernel size 92 1.3.2.2 pgoyette * %r8 : loaded start address 93 1.3.2.2 pgoyette * %r9 : kernel entry address 94 1.3.2.2 pgoyette */ 95 1.3.2.2 pgoyette 96 1.3.2.2 pgoyette cld /* LynxOS depends on it */ 97 1.3.2.2 pgoyette 98 1.3.2.2 pgoyette cli 99 1.3.2.2 pgoyette 100 1.3.2.2 pgoyette /* Copy kernel */ 101 1.3.2.2 pgoyette mov %rcx, %r12 /* original kernel size */ 102 1.3.2.2 pgoyette movq %rdi, %r11 /* for misaligned check */ 103 1.3.2.2 pgoyette 104 1.3.2.2 pgoyette #if !defined(NO_OVERLAP) 105 1.3.2.2 pgoyette movq %rdi, %r13 106 1.3.2.2 pgoyette subq %rsi, %r13 107 1.3.2.2 pgoyette #endif 108 1.3.2.2 pgoyette 109 1.3.2.2 pgoyette shrq $3, %rcx /* count for copy by words */ 110 1.3.2.2 pgoyette jz 8f /* j if less than 8 bytes */ 111 1.3.2.2 pgoyette 112 1.3.2.2 pgoyette lea -8(%rdi, %r12), %r14 /* target address of last 8 */ 113 1.3.2.2 pgoyette mov -8(%rsi, %r12), %r15 /* get last word */ 114 1.3.2.2 pgoyette #if !defined(NO_OVERLAP) 115 1.3.2.2 pgoyette cmpq %r12, %r13 /* overlapping? */ 116 1.3.2.2 pgoyette jb 10f 117 1.3.2.2 pgoyette #endif 118 1.3.2.2 pgoyette 119 1.3.2.2 pgoyette /* 120 1.3.2.2 pgoyette * Non-overlaping, copy forwards. 121 1.3.2.2 pgoyette * Newer Intel cpus (Nehalem) will do 16byte read/write transfers 122 1.3.2.2 pgoyette * if %ecx is more than 76. 123 1.3.2.2 pgoyette * AMD might do something similar some day. 124 1.3.2.2 pgoyette */ 125 1.3.2.2 pgoyette and $7, %r11 /* destination misaligned ? */ 126 1.3.2.2 pgoyette jnz 2f 127 1.3.2.2 pgoyette rep 128 1.3.2.2 pgoyette movsq 129 1.3.2.2 pgoyette mov %r15, (%r14) /* write last word */ 130 1.3.2.2 pgoyette jmp .Lcopy_done 131 1.3.2.2 pgoyette 132 1.3.2.2 pgoyette /* 133 1.3.2.2 pgoyette * Destination misaligned 134 1.3.2.2 pgoyette * AMD say it is better to align the destination (not the source). 135 1.3.2.2 pgoyette * This will also re-align copies if the source and dest are both 136 1.3.2.2 pgoyette * misaligned by the same amount) 137 1.3.2.2 pgoyette * (I think Nehalem will use its accelerated copy if the source 138 1.3.2.2 pgoyette * and destination have the same alignment.) 139 1.3.2.2 pgoyette */ 140 1.3.2.2 pgoyette 2: 141 1.3.2.2 pgoyette lea -9(%r11, %r12), %rcx /* post re-alignment count */ 142 1.3.2.2 pgoyette neg %r11 /* now -1 .. -7 */ 143 1.3.2.2 pgoyette mov (%rsi), %r12 /* get first word */ 144 1.3.2.2 pgoyette mov %rdi, %r13 /* target for first word */ 145 1.3.2.2 pgoyette lea 8(%rsi, %r11), %rsi 146 1.3.2.2 pgoyette lea 8(%rdi, %r11), %rdi 147 1.3.2.2 pgoyette shr $3, %rcx 148 1.3.2.2 pgoyette rep 149 1.3.2.2 pgoyette movsq 150 1.3.2.2 pgoyette mov %r12, (%r13) /* write first word */ 151 1.3.2.2 pgoyette mov %r15, (%r14) /* write last word */ 152 1.3.2.2 pgoyette jmp .Lcopy_done 153 1.3.2.2 pgoyette 154 1.3.2.2 pgoyette #if !defined(NO_OVERLAP) 155 1.3.2.2 pgoyette /* Must copy backwards. 156 1.3.2.2 pgoyette * Reverse copy is probably easy to code faster than 'rep movds' 157 1.3.2.2 pgoyette * since that requires (IIRC) an extra clock every 3 iterations (AMD). 158 1.3.2.2 pgoyette * However I don't suppose anything cares that much! 159 1.3.2.2 pgoyette * The big cost is the std/cld pair - reputedly 50+ cycles on Netburst P4. 160 1.3.2.2 pgoyette * The copy is aligned with the buffer start (more likely to 161 1.3.2.2 pgoyette * be a multiple of 8 than the end). 162 1.3.2.2 pgoyette */ 163 1.3.2.2 pgoyette 10: 164 1.3.2.2 pgoyette lea -8(%rsi, %rcx, 8), %rsi 165 1.3.2.2 pgoyette lea -8(%rdi, %rcx, 8), %rdi 166 1.3.2.2 pgoyette std 167 1.3.2.2 pgoyette rep 168 1.3.2.2 pgoyette movsq 169 1.3.2.2 pgoyette cld 170 1.3.2.2 pgoyette mov %r15, (%r14) /* write last bytes */ 171 1.3.2.2 pgoyette jmp .Lcopy_done 172 1.3.2.2 pgoyette #endif 173 1.3.2.2 pgoyette 174 1.3.2.2 pgoyette /* Less than 8 bytes to copy, copy by bytes */ 175 1.3.2.2 pgoyette /* Intel Nehalem optimise 'rep movsb' for <= 7 bytes (9-15 clocks). 176 1.3.2.2 pgoyette * For longer transfers it is 50+ ! 177 1.3.2.2 pgoyette */ 178 1.3.2.2 pgoyette 8: mov %r12, %rcx 179 1.3.2.2 pgoyette 180 1.3.2.2 pgoyette #if !defined(NO_OVERLAP) 181 1.3.2.2 pgoyette cmpq %r12, %r13 /* overlapping? */ 182 1.3.2.2 pgoyette jb 81f 183 1.3.2.2 pgoyette #endif 184 1.3.2.2 pgoyette 185 1.3.2.2 pgoyette /* nope, copy forwards. */ 186 1.3.2.2 pgoyette rep 187 1.3.2.2 pgoyette movsb 188 1.3.2.2 pgoyette jmp .Lcopy_done 189 1.3.2.2 pgoyette 190 1.3.2.2 pgoyette #if !defined(NO_OVERLAP) 191 1.3.2.2 pgoyette /* Must copy backwards */ 192 1.3.2.2 pgoyette 81: 193 1.3.2.2 pgoyette lea -1(%rsi, %rcx), %rsi 194 1.3.2.2 pgoyette lea -1(%rdi, %rcx), %rdi 195 1.3.2.2 pgoyette std 196 1.3.2.2 pgoyette rep 197 1.3.2.2 pgoyette movsb 198 1.3.2.2 pgoyette cld 199 1.3.2.2 pgoyette #endif 200 1.3.2.2 pgoyette /* End of copy kernel */ 201 1.3.2.2 pgoyette .Lcopy_done: 202 1.3.2.2 pgoyette 203 1.3.2.2 pgoyette mov %r8, %rdi /* %rdi: loaded start address */ 204 1.3.2.2 pgoyette mov %r9, %rsi /* %rsi: kernel entry address */ 205 1.3.2.2 pgoyette 206 1.3.2.2 pgoyette /* Prepare jump address */ 207 1.3.2.2 pgoyette lea (start32a - start)(%rdi), %rax 208 1.3.2.2 pgoyette movl %eax, (start32r - start)(%rdi) 209 1.3.2.2 pgoyette 210 1.3.2.2 pgoyette /* Setup GDT */ 211 1.3.2.2 pgoyette lea (gdt - start)(%rdi), %rax 212 1.3.2.2 pgoyette mov %rax, (gdtrr - start)(%rdi) 213 1.3.2.2 pgoyette lgdt (gdtr - start)(%rdi) 214 1.3.2.2 pgoyette 215 1.3.2.2 pgoyette /* Jump to set %cs */ 216 1.3.2.2 pgoyette ljmp *(start32r - start)(%rdi) 217 1.3.2.2 pgoyette 218 1.3.2.2 pgoyette .align 4 219 1.3.2.2 pgoyette .code32 220 1.3.2.2 pgoyette start32a: 221 1.3.2.2 pgoyette movl $DATA_SEGMENT, %eax 222 1.3.2.2 pgoyette movw %ax, %ds 223 1.3.2.2 pgoyette movw %ax, %es 224 1.3.2.2 pgoyette movw %ax, %fs 225 1.3.2.2 pgoyette movw %ax, %gs 226 1.3.2.2 pgoyette movw %ax, %ss 227 1.3.2.2 pgoyette 228 1.3.2.2 pgoyette movl %edx, %esp 229 1.3.2.2 pgoyette 230 1.3.2.2 pgoyette /* Disable Paging in CR0 */ 231 1.3.2.2 pgoyette movl %cr0, %eax 232 1.3.2.2 pgoyette andl $(~CR0_PG), %eax 233 1.3.2.2 pgoyette movl %eax, %cr0 234 1.3.2.2 pgoyette 235 1.3.2.2 pgoyette /* Disable PAE in CR4 */ 236 1.3.2.2 pgoyette movl %cr4, %eax 237 1.3.2.2 pgoyette andl $(~CR4_PAE), %eax 238 1.3.2.2 pgoyette movl %eax, %cr4 239 1.3.2.2 pgoyette 240 1.3.2.2 pgoyette jmp start32b 241 1.3.2.2 pgoyette 242 1.3.2.2 pgoyette .align 4 243 1.3.2.2 pgoyette start32b: 244 1.3.2.2 pgoyette xor %eax, %eax 245 1.3.2.2 pgoyette call *%esi 246 1.3.2.2 pgoyette 247 1.3.2.2 pgoyette .align 16 248 1.3.2.2 pgoyette start32r: 249 1.3.2.2 pgoyette .long 0 250 1.3.2.2 pgoyette .long CODE_SEGMENT 251 1.3.2.2 pgoyette .align 16 252 1.3.2.2 pgoyette gdt: 253 1.3.2.2 pgoyette .long 0, 0 254 1.3.2.2 pgoyette .byte 0xff, 0xff, 0x00, 0x00, 0x00, 0x9f, 0xcf, 0x00 255 1.3.2.2 pgoyette .byte 0xff, 0xff, 0x00, 0x00, 0x00, 0x93, 0xcf, 0x00 256 1.3.2.2 pgoyette gdtr: 257 1.3.2.2 pgoyette .word gdtr - gdt 258 1.3.2.2 pgoyette gdtrr: 259 1.3.2.2 pgoyette .quad 260 1.3.2.2 pgoyette start32end: 261 1.3.2.2 pgoyette /* Space for the stack */ 262 1.3.2.2 pgoyette .align 16 263 1.3.2.2 pgoyette .space 8192 264 1.3.2.2 pgoyette startprog64_end: 265
Indexes created Sat Oct 25 10:09:55 GMT 2025