root / tests / test-i386.c @ 776f2227
History | View | Annotate | Download (63.3 kB)
| 1 |
/*
|
|---|---|
| 2 |
* x86 CPU test
|
| 3 |
*
|
| 4 |
* Copyright (c) 2003 Fabrice Bellard
|
| 5 |
*
|
| 6 |
* This program is free software; you can redistribute it and/or modify
|
| 7 |
* it under the terms of the GNU General Public License as published by
|
| 8 |
* the Free Software Foundation; either version 2 of the License, or
|
| 9 |
* (at your option) any later version.
|
| 10 |
*
|
| 11 |
* This program is distributed in the hope that it will be useful,
|
| 12 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
| 13 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
| 14 |
* GNU General Public License for more details.
|
| 15 |
*
|
| 16 |
* You should have received a copy of the GNU General Public License
|
| 17 |
* along with this program; if not, write to the Free Software
|
| 18 |
* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
|
| 19 |
*/
|
| 20 |
#define _GNU_SOURCE
|
| 21 |
#include <stdlib.h> |
| 22 |
#include <stdio.h> |
| 23 |
#include <string.h> |
| 24 |
#include <inttypes.h> |
| 25 |
#include <math.h> |
| 26 |
#include <signal.h> |
| 27 |
#include <setjmp.h> |
| 28 |
#include <errno.h> |
| 29 |
#include <sys/ucontext.h> |
| 30 |
#include <sys/mman.h> |
| 31 |
|
| 32 |
#if !defined(__x86_64__)
|
| 33 |
#define TEST_VM86
|
| 34 |
#define TEST_SEGS
|
| 35 |
#endif
|
| 36 |
//#define LINUX_VM86_IOPL_FIX
|
| 37 |
//#define TEST_P4_FLAGS
|
| 38 |
#if defined(__x86_64__)
|
| 39 |
#define TEST_SSE
|
| 40 |
#define TEST_CMOV 1 |
| 41 |
#define TEST_FCOMI 1 |
| 42 |
#else
|
| 43 |
#define TEST_CMOV 0 |
| 44 |
#define TEST_FCOMI 0 |
| 45 |
#endif
|
| 46 |
|
| 47 |
#if defined(__x86_64__)
|
| 48 |
#define FMT64X "%016lx" |
| 49 |
#define FMTLX "%016lx" |
| 50 |
#define X86_64_ONLY(x) x
|
| 51 |
#else
|
| 52 |
#define FMT64X "%016llx" |
| 53 |
#define FMTLX "%08lx" |
| 54 |
#define X86_64_ONLY(x)
|
| 55 |
#endif
|
| 56 |
|
| 57 |
#ifdef TEST_VM86
|
| 58 |
#include <asm/vm86.h> |
| 59 |
#endif
|
| 60 |
|
| 61 |
#define xglue(x, y) x ## y |
| 62 |
#define glue(x, y) xglue(x, y)
|
| 63 |
#define stringify(s) tostring(s)
|
| 64 |
#define tostring(s) #s |
| 65 |
|
| 66 |
#define CC_C 0x0001 |
| 67 |
#define CC_P 0x0004 |
| 68 |
#define CC_A 0x0010 |
| 69 |
#define CC_Z 0x0040 |
| 70 |
#define CC_S 0x0080 |
| 71 |
#define CC_O 0x0800 |
| 72 |
|
| 73 |
#define __init_call __attribute__ ((unused,__section__ ("initcall"))) |
| 74 |
|
| 75 |
#define CC_MASK (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A)
|
| 76 |
|
| 77 |
#if defined(__x86_64__)
|
| 78 |
static inline long i2l(long v) |
| 79 |
{
|
| 80 |
return v | ((v ^ 0xabcd) << 32); |
| 81 |
} |
| 82 |
#else
|
| 83 |
static inline long i2l(long v) |
| 84 |
{
|
| 85 |
return v;
|
| 86 |
} |
| 87 |
#endif
|
| 88 |
|
| 89 |
#define OP add
|
| 90 |
#include "test-i386.h" |
| 91 |
|
| 92 |
#define OP sub
|
| 93 |
#include "test-i386.h" |
| 94 |
|
| 95 |
#define OP xor
|
| 96 |
#include "test-i386.h" |
| 97 |
|
| 98 |
#define OP and
|
| 99 |
#include "test-i386.h" |
| 100 |
|
| 101 |
#define OP or
|
| 102 |
#include "test-i386.h" |
| 103 |
|
| 104 |
#define OP cmp
|
| 105 |
#include "test-i386.h" |
| 106 |
|
| 107 |
#define OP adc
|
| 108 |
#define OP_CC
|
| 109 |
#include "test-i386.h" |
| 110 |
|
| 111 |
#define OP sbb
|
| 112 |
#define OP_CC
|
| 113 |
#include "test-i386.h" |
| 114 |
|
| 115 |
#define OP inc
|
| 116 |
#define OP_CC
|
| 117 |
#define OP1
|
| 118 |
#include "test-i386.h" |
| 119 |
|
| 120 |
#define OP dec
|
| 121 |
#define OP_CC
|
| 122 |
#define OP1
|
| 123 |
#include "test-i386.h" |
| 124 |
|
| 125 |
#define OP neg
|
| 126 |
#define OP_CC
|
| 127 |
#define OP1
|
| 128 |
#include "test-i386.h" |
| 129 |
|
| 130 |
#define OP not
|
| 131 |
#define OP_CC
|
| 132 |
#define OP1
|
| 133 |
#include "test-i386.h" |
| 134 |
|
| 135 |
#undef CC_MASK
|
| 136 |
#define CC_MASK (CC_C | CC_P | CC_Z | CC_S | CC_O)
|
| 137 |
|
| 138 |
#define OP shl
|
| 139 |
#include "test-i386-shift.h" |
| 140 |
|
| 141 |
#define OP shr
|
| 142 |
#include "test-i386-shift.h" |
| 143 |
|
| 144 |
#define OP sar
|
| 145 |
#include "test-i386-shift.h" |
| 146 |
|
| 147 |
#define OP rol
|
| 148 |
#include "test-i386-shift.h" |
| 149 |
|
| 150 |
#define OP ror
|
| 151 |
#include "test-i386-shift.h" |
| 152 |
|
| 153 |
#define OP rcr
|
| 154 |
#define OP_CC
|
| 155 |
#include "test-i386-shift.h" |
| 156 |
|
| 157 |
#define OP rcl
|
| 158 |
#define OP_CC
|
| 159 |
#include "test-i386-shift.h" |
| 160 |
|
| 161 |
#define OP shld
|
| 162 |
#define OP_SHIFTD
|
| 163 |
#define OP_NOBYTE
|
| 164 |
#include "test-i386-shift.h" |
| 165 |
|
| 166 |
#define OP shrd
|
| 167 |
#define OP_SHIFTD
|
| 168 |
#define OP_NOBYTE
|
| 169 |
#include "test-i386-shift.h" |
| 170 |
|
| 171 |
/* XXX: should be more precise ? */
|
| 172 |
#undef CC_MASK
|
| 173 |
#define CC_MASK (CC_C)
|
| 174 |
|
| 175 |
#define OP bt
|
| 176 |
#define OP_NOBYTE
|
| 177 |
#include "test-i386-shift.h" |
| 178 |
|
| 179 |
#define OP bts
|
| 180 |
#define OP_NOBYTE
|
| 181 |
#include "test-i386-shift.h" |
| 182 |
|
| 183 |
#define OP btr
|
| 184 |
#define OP_NOBYTE
|
| 185 |
#include "test-i386-shift.h" |
| 186 |
|
| 187 |
#define OP btc
|
| 188 |
#define OP_NOBYTE
|
| 189 |
#include "test-i386-shift.h" |
| 190 |
|
| 191 |
/* lea test (modrm support) */
|
| 192 |
#define TEST_LEAQ(STR)\
|
| 193 |
{\
|
| 194 |
asm("lea " STR ", %0"\ |
| 195 |
: "=r" (res)\
|
| 196 |
: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\ |
| 197 |
printf("lea %s = " FMTLX "\n", STR, res);\ |
| 198 |
} |
| 199 |
|
| 200 |
#define TEST_LEA(STR)\
|
| 201 |
{\
|
| 202 |
asm("lea " STR ", %0"\ |
| 203 |
: "=r" (res)\
|
| 204 |
: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\ |
| 205 |
printf("lea %s = " FMTLX "\n", STR, res);\ |
| 206 |
} |
| 207 |
|
| 208 |
#define TEST_LEA16(STR)\
|
| 209 |
{\
|
| 210 |
asm(".code16 ; .byte 0x67 ; leal " STR ", %0 ; .code32"\ |
| 211 |
: "=wq" (res)\
|
| 212 |
: "a" (eax), "b" (ebx), "c" (ecx), "d" (edx), "S" (esi), "D" (edi));\ |
| 213 |
printf("lea %s = %08lx\n", STR, res);\
|
| 214 |
} |
| 215 |
|
| 216 |
|
| 217 |
void test_lea(void) |
| 218 |
{
|
| 219 |
long eax, ebx, ecx, edx, esi, edi, res;
|
| 220 |
eax = i2l(0x0001);
|
| 221 |
ebx = i2l(0x0002);
|
| 222 |
ecx = i2l(0x0004);
|
| 223 |
edx = i2l(0x0008);
|
| 224 |
esi = i2l(0x0010);
|
| 225 |
edi = i2l(0x0020);
|
| 226 |
|
| 227 |
TEST_LEA("0x4000");
|
| 228 |
|
| 229 |
TEST_LEA("(%%eax)");
|
| 230 |
TEST_LEA("(%%ebx)");
|
| 231 |
TEST_LEA("(%%ecx)");
|
| 232 |
TEST_LEA("(%%edx)");
|
| 233 |
TEST_LEA("(%%esi)");
|
| 234 |
TEST_LEA("(%%edi)");
|
| 235 |
|
| 236 |
TEST_LEA("0x40(%%eax)");
|
| 237 |
TEST_LEA("0x40(%%ebx)");
|
| 238 |
TEST_LEA("0x40(%%ecx)");
|
| 239 |
TEST_LEA("0x40(%%edx)");
|
| 240 |
TEST_LEA("0x40(%%esi)");
|
| 241 |
TEST_LEA("0x40(%%edi)");
|
| 242 |
|
| 243 |
TEST_LEA("0x4000(%%eax)");
|
| 244 |
TEST_LEA("0x4000(%%ebx)");
|
| 245 |
TEST_LEA("0x4000(%%ecx)");
|
| 246 |
TEST_LEA("0x4000(%%edx)");
|
| 247 |
TEST_LEA("0x4000(%%esi)");
|
| 248 |
TEST_LEA("0x4000(%%edi)");
|
| 249 |
|
| 250 |
TEST_LEA("(%%eax, %%ecx)");
|
| 251 |
TEST_LEA("(%%ebx, %%edx)");
|
| 252 |
TEST_LEA("(%%ecx, %%ecx)");
|
| 253 |
TEST_LEA("(%%edx, %%ecx)");
|
| 254 |
TEST_LEA("(%%esi, %%ecx)");
|
| 255 |
TEST_LEA("(%%edi, %%ecx)");
|
| 256 |
|
| 257 |
TEST_LEA("0x40(%%eax, %%ecx)");
|
| 258 |
TEST_LEA("0x4000(%%ebx, %%edx)");
|
| 259 |
|
| 260 |
TEST_LEA("(%%ecx, %%ecx, 2)");
|
| 261 |
TEST_LEA("(%%edx, %%ecx, 4)");
|
| 262 |
TEST_LEA("(%%esi, %%ecx, 8)");
|
| 263 |
|
| 264 |
TEST_LEA("(,%%eax, 2)");
|
| 265 |
TEST_LEA("(,%%ebx, 4)");
|
| 266 |
TEST_LEA("(,%%ecx, 8)");
|
| 267 |
|
| 268 |
TEST_LEA("0x40(,%%eax, 2)");
|
| 269 |
TEST_LEA("0x40(,%%ebx, 4)");
|
| 270 |
TEST_LEA("0x40(,%%ecx, 8)");
|
| 271 |
|
| 272 |
|
| 273 |
TEST_LEA("-10(%%ecx, %%ecx, 2)");
|
| 274 |
TEST_LEA("-10(%%edx, %%ecx, 4)");
|
| 275 |
TEST_LEA("-10(%%esi, %%ecx, 8)");
|
| 276 |
|
| 277 |
TEST_LEA("0x4000(%%ecx, %%ecx, 2)");
|
| 278 |
TEST_LEA("0x4000(%%edx, %%ecx, 4)");
|
| 279 |
TEST_LEA("0x4000(%%esi, %%ecx, 8)");
|
| 280 |
|
| 281 |
#if defined(__x86_64__)
|
| 282 |
TEST_LEAQ("0x4000");
|
| 283 |
TEST_LEAQ("0x4000(%%rip)");
|
| 284 |
|
| 285 |
TEST_LEAQ("(%%rax)");
|
| 286 |
TEST_LEAQ("(%%rbx)");
|
| 287 |
TEST_LEAQ("(%%rcx)");
|
| 288 |
TEST_LEAQ("(%%rdx)");
|
| 289 |
TEST_LEAQ("(%%rsi)");
|
| 290 |
TEST_LEAQ("(%%rdi)");
|
| 291 |
|
| 292 |
TEST_LEAQ("0x40(%%rax)");
|
| 293 |
TEST_LEAQ("0x40(%%rbx)");
|
| 294 |
TEST_LEAQ("0x40(%%rcx)");
|
| 295 |
TEST_LEAQ("0x40(%%rdx)");
|
| 296 |
TEST_LEAQ("0x40(%%rsi)");
|
| 297 |
TEST_LEAQ("0x40(%%rdi)");
|
| 298 |
|
| 299 |
TEST_LEAQ("0x4000(%%rax)");
|
| 300 |
TEST_LEAQ("0x4000(%%rbx)");
|
| 301 |
TEST_LEAQ("0x4000(%%rcx)");
|
| 302 |
TEST_LEAQ("0x4000(%%rdx)");
|
| 303 |
TEST_LEAQ("0x4000(%%rsi)");
|
| 304 |
TEST_LEAQ("0x4000(%%rdi)");
|
| 305 |
|
| 306 |
TEST_LEAQ("(%%rax, %%rcx)");
|
| 307 |
TEST_LEAQ("(%%rbx, %%rdx)");
|
| 308 |
TEST_LEAQ("(%%rcx, %%rcx)");
|
| 309 |
TEST_LEAQ("(%%rdx, %%rcx)");
|
| 310 |
TEST_LEAQ("(%%rsi, %%rcx)");
|
| 311 |
TEST_LEAQ("(%%rdi, %%rcx)");
|
| 312 |
|
| 313 |
TEST_LEAQ("0x40(%%rax, %%rcx)");
|
| 314 |
TEST_LEAQ("0x4000(%%rbx, %%rdx)");
|
| 315 |
|
| 316 |
TEST_LEAQ("(%%rcx, %%rcx, 2)");
|
| 317 |
TEST_LEAQ("(%%rdx, %%rcx, 4)");
|
| 318 |
TEST_LEAQ("(%%rsi, %%rcx, 8)");
|
| 319 |
|
| 320 |
TEST_LEAQ("(,%%rax, 2)");
|
| 321 |
TEST_LEAQ("(,%%rbx, 4)");
|
| 322 |
TEST_LEAQ("(,%%rcx, 8)");
|
| 323 |
|
| 324 |
TEST_LEAQ("0x40(,%%rax, 2)");
|
| 325 |
TEST_LEAQ("0x40(,%%rbx, 4)");
|
| 326 |
TEST_LEAQ("0x40(,%%rcx, 8)");
|
| 327 |
|
| 328 |
|
| 329 |
TEST_LEAQ("-10(%%rcx, %%rcx, 2)");
|
| 330 |
TEST_LEAQ("-10(%%rdx, %%rcx, 4)");
|
| 331 |
TEST_LEAQ("-10(%%rsi, %%rcx, 8)");
|
| 332 |
|
| 333 |
TEST_LEAQ("0x4000(%%rcx, %%rcx, 2)");
|
| 334 |
TEST_LEAQ("0x4000(%%rdx, %%rcx, 4)");
|
| 335 |
TEST_LEAQ("0x4000(%%rsi, %%rcx, 8)");
|
| 336 |
#else
|
| 337 |
/* limited 16 bit addressing test */
|
| 338 |
TEST_LEA16("0x4000");
|
| 339 |
TEST_LEA16("(%%bx)");
|
| 340 |
TEST_LEA16("(%%si)");
|
| 341 |
TEST_LEA16("(%%di)");
|
| 342 |
TEST_LEA16("0x40(%%bx)");
|
| 343 |
TEST_LEA16("0x40(%%si)");
|
| 344 |
TEST_LEA16("0x40(%%di)");
|
| 345 |
TEST_LEA16("0x4000(%%bx)");
|
| 346 |
TEST_LEA16("0x4000(%%si)");
|
| 347 |
TEST_LEA16("(%%bx,%%si)");
|
| 348 |
TEST_LEA16("(%%bx,%%di)");
|
| 349 |
TEST_LEA16("0x40(%%bx,%%si)");
|
| 350 |
TEST_LEA16("0x40(%%bx,%%di)");
|
| 351 |
TEST_LEA16("0x4000(%%bx,%%si)");
|
| 352 |
TEST_LEA16("0x4000(%%bx,%%di)");
|
| 353 |
#endif
|
| 354 |
} |
| 355 |
|
| 356 |
#define TEST_JCC(JCC, v1, v2)\
|
| 357 |
{\
|
| 358 |
int res;\
|
| 359 |
asm("movl $1, %0\n\t"\ |
| 360 |
"cmpl %2, %1\n\t"\
|
| 361 |
"j" JCC " 1f\n\t"\ |
| 362 |
"movl $0, %0\n\t"\
|
| 363 |
"1:\n\t"\
|
| 364 |
: "=r" (res)\
|
| 365 |
: "r" (v1), "r" (v2));\ |
| 366 |
printf("%-10s %d\n", "j" JCC, res);\ |
| 367 |
\ |
| 368 |
asm("movl $0, %0\n\t"\ |
| 369 |
"cmpl %2, %1\n\t"\
|
| 370 |
"set" JCC " %b0\n\t"\ |
| 371 |
: "=r" (res)\
|
| 372 |
: "r" (v1), "r" (v2));\ |
| 373 |
printf("%-10s %d\n", "set" JCC, res);\ |
| 374 |
if (TEST_CMOV) {\
|
| 375 |
long val = i2l(1);\ |
| 376 |
long res = i2l(0x12345678);\ |
| 377 |
X86_64_ONLY(\ |
| 378 |
asm("cmpl %2, %1\n\t"\ |
| 379 |
"cmov" JCC "q %3, %0\n\t"\ |
| 380 |
: "=r" (res)\
|
| 381 |
: "r" (v1), "r" (v2), "m" (val), "0" (res));\ |
| 382 |
printf("%-10s R=" FMTLX "\n", "cmov" JCC "l", res);)\ |
| 383 |
asm("cmpl %2, %1\n\t"\ |
| 384 |
"cmov" JCC "l %k3, %k0\n\t"\ |
| 385 |
: "=r" (res)\
|
| 386 |
: "r" (v1), "r" (v2), "m" (val), "0" (res));\ |
| 387 |
printf("%-10s R=" FMTLX "\n", "cmov" JCC "l", res);\ |
| 388 |
asm("cmpl %2, %1\n\t"\ |
| 389 |
"cmov" JCC "w %w3, %w0\n\t"\ |
| 390 |
: "=r" (res)\
|
| 391 |
: "r" (v1), "r" (v2), "r" (1), "0" (res));\ |
| 392 |
printf("%-10s R=" FMTLX "\n", "cmov" JCC "w", res);\ |
| 393 |
} \ |
| 394 |
} |
| 395 |
|
| 396 |
/* various jump tests */
|
| 397 |
void test_jcc(void) |
| 398 |
{
|
| 399 |
TEST_JCC("ne", 1, 1); |
| 400 |
TEST_JCC("ne", 1, 0); |
| 401 |
|
| 402 |
TEST_JCC("e", 1, 1); |
| 403 |
TEST_JCC("e", 1, 0); |
| 404 |
|
| 405 |
TEST_JCC("l", 1, 1); |
| 406 |
TEST_JCC("l", 1, 0); |
| 407 |
TEST_JCC("l", 1, -1); |
| 408 |
|
| 409 |
TEST_JCC("le", 1, 1); |
| 410 |
TEST_JCC("le", 1, 0); |
| 411 |
TEST_JCC("le", 1, -1); |
| 412 |
|
| 413 |
TEST_JCC("ge", 1, 1); |
| 414 |
TEST_JCC("ge", 1, 0); |
| 415 |
TEST_JCC("ge", -1, 1); |
| 416 |
|
| 417 |
TEST_JCC("g", 1, 1); |
| 418 |
TEST_JCC("g", 1, 0); |
| 419 |
TEST_JCC("g", 1, -1); |
| 420 |
|
| 421 |
TEST_JCC("b", 1, 1); |
| 422 |
TEST_JCC("b", 1, 0); |
| 423 |
TEST_JCC("b", 1, -1); |
| 424 |
|
| 425 |
TEST_JCC("be", 1, 1); |
| 426 |
TEST_JCC("be", 1, 0); |
| 427 |
TEST_JCC("be", 1, -1); |
| 428 |
|
| 429 |
TEST_JCC("ae", 1, 1); |
| 430 |
TEST_JCC("ae", 1, 0); |
| 431 |
TEST_JCC("ae", 1, -1); |
| 432 |
|
| 433 |
TEST_JCC("a", 1, 1); |
| 434 |
TEST_JCC("a", 1, 0); |
| 435 |
TEST_JCC("a", 1, -1); |
| 436 |
|
| 437 |
|
| 438 |
TEST_JCC("p", 1, 1); |
| 439 |
TEST_JCC("p", 1, 0); |
| 440 |
|
| 441 |
TEST_JCC("np", 1, 1); |
| 442 |
TEST_JCC("np", 1, 0); |
| 443 |
|
| 444 |
TEST_JCC("o", 0x7fffffff, 0); |
| 445 |
TEST_JCC("o", 0x7fffffff, -1); |
| 446 |
|
| 447 |
TEST_JCC("no", 0x7fffffff, 0); |
| 448 |
TEST_JCC("no", 0x7fffffff, -1); |
| 449 |
|
| 450 |
TEST_JCC("s", 0, 1); |
| 451 |
TEST_JCC("s", 0, -1); |
| 452 |
TEST_JCC("s", 0, 0); |
| 453 |
|
| 454 |
TEST_JCC("ns", 0, 1); |
| 455 |
TEST_JCC("ns", 0, -1); |
| 456 |
TEST_JCC("ns", 0, 0); |
| 457 |
} |
| 458 |
|
| 459 |
#undef CC_MASK
|
| 460 |
#ifdef TEST_P4_FLAGS
|
| 461 |
#define CC_MASK (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A)
|
| 462 |
#else
|
| 463 |
#define CC_MASK (CC_O | CC_C)
|
| 464 |
#endif
|
| 465 |
|
| 466 |
#define OP mul
|
| 467 |
#include "test-i386-muldiv.h" |
| 468 |
|
| 469 |
#define OP imul
|
| 470 |
#include "test-i386-muldiv.h" |
| 471 |
|
| 472 |
void test_imulw2(long op0, long op1) |
| 473 |
{
|
| 474 |
long res, s1, s0, flags;
|
| 475 |
s0 = op0; |
| 476 |
s1 = op1; |
| 477 |
res = s0; |
| 478 |
flags = 0;
|
| 479 |
asm volatile ("push %4\n\t" |
| 480 |
"popf\n\t"
|
| 481 |
"imulw %w2, %w0\n\t"
|
| 482 |
"pushf\n\t"
|
| 483 |
"pop %1\n\t"
|
| 484 |
: "=q" (res), "=g" (flags) |
| 485 |
: "q" (s1), "0" (res), "1" (flags)); |
| 486 |
printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n", |
| 487 |
"imulw", s0, s1, res, flags & CC_MASK);
|
| 488 |
} |
| 489 |
|
| 490 |
void test_imull2(long op0, long op1) |
| 491 |
{
|
| 492 |
long res, s1, s0, flags;
|
| 493 |
s0 = op0; |
| 494 |
s1 = op1; |
| 495 |
res = s0; |
| 496 |
flags = 0;
|
| 497 |
asm volatile ("push %4\n\t" |
| 498 |
"popf\n\t"
|
| 499 |
"imull %k2, %k0\n\t"
|
| 500 |
"pushf\n\t"
|
| 501 |
"pop %1\n\t"
|
| 502 |
: "=q" (res), "=g" (flags) |
| 503 |
: "q" (s1), "0" (res), "1" (flags)); |
| 504 |
printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n", |
| 505 |
"imull", s0, s1, res, flags & CC_MASK);
|
| 506 |
} |
| 507 |
|
| 508 |
#if defined(__x86_64__)
|
| 509 |
void test_imulq2(long op0, long op1) |
| 510 |
{
|
| 511 |
long res, s1, s0, flags;
|
| 512 |
s0 = op0; |
| 513 |
s1 = op1; |
| 514 |
res = s0; |
| 515 |
flags = 0;
|
| 516 |
asm volatile ("push %4\n\t" |
| 517 |
"popf\n\t"
|
| 518 |
"imulq %2, %0\n\t"
|
| 519 |
"pushf\n\t"
|
| 520 |
"pop %1\n\t"
|
| 521 |
: "=q" (res), "=g" (flags) |
| 522 |
: "q" (s1), "0" (res), "1" (flags)); |
| 523 |
printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n", |
| 524 |
"imulq", s0, s1, res, flags & CC_MASK);
|
| 525 |
} |
| 526 |
#endif
|
| 527 |
|
| 528 |
#define TEST_IMUL_IM(size, rsize, op0, op1)\
|
| 529 |
{\
|
| 530 |
long res, flags, s1;\
|
| 531 |
flags = 0;\
|
| 532 |
res = 0;\
|
| 533 |
s1 = op1;\ |
| 534 |
asm volatile ("push %3\n\t"\ |
| 535 |
"popf\n\t"\
|
| 536 |
"imul" size " $" #op0 ", %" rsize "2, %" rsize "0\n\t" \ |
| 537 |
"pushf\n\t"\
|
| 538 |
"pop %1\n\t"\
|
| 539 |
: "=r" (res), "=g" (flags)\ |
| 540 |
: "r" (s1), "1" (flags), "0" (res));\ |
| 541 |
printf("%-10s A=" FMTLX " B=" FMTLX " R=" FMTLX " CC=%04lx\n",\ |
| 542 |
"imul" size " im", (long)op0, (long)op1, res, flags & CC_MASK);\ |
| 543 |
} |
| 544 |
|
| 545 |
|
| 546 |
#undef CC_MASK
|
| 547 |
#define CC_MASK (0) |
| 548 |
|
| 549 |
#define OP div
|
| 550 |
#include "test-i386-muldiv.h" |
| 551 |
|
| 552 |
#define OP idiv
|
| 553 |
#include "test-i386-muldiv.h" |
| 554 |
|
| 555 |
void test_mul(void) |
| 556 |
{
|
| 557 |
test_imulb(0x1234561d, 4); |
| 558 |
test_imulb(3, -4); |
| 559 |
test_imulb(0x80, 0x80); |
| 560 |
test_imulb(0x10, 0x10); |
| 561 |
|
| 562 |
test_imulw(0, 0x1234001d, 45); |
| 563 |
test_imulw(0, 23, -45); |
| 564 |
test_imulw(0, 0x8000, 0x8000); |
| 565 |
test_imulw(0, 0x100, 0x100); |
| 566 |
|
| 567 |
test_imull(0, 0x1234001d, 45); |
| 568 |
test_imull(0, 23, -45); |
| 569 |
test_imull(0, 0x80000000, 0x80000000); |
| 570 |
test_imull(0, 0x10000, 0x10000); |
| 571 |
|
| 572 |
test_mulb(0x1234561d, 4); |
| 573 |
test_mulb(3, -4); |
| 574 |
test_mulb(0x80, 0x80); |
| 575 |
test_mulb(0x10, 0x10); |
| 576 |
|
| 577 |
test_mulw(0, 0x1234001d, 45); |
| 578 |
test_mulw(0, 23, -45); |
| 579 |
test_mulw(0, 0x8000, 0x8000); |
| 580 |
test_mulw(0, 0x100, 0x100); |
| 581 |
|
| 582 |
test_mull(0, 0x1234001d, 45); |
| 583 |
test_mull(0, 23, -45); |
| 584 |
test_mull(0, 0x80000000, 0x80000000); |
| 585 |
test_mull(0, 0x10000, 0x10000); |
| 586 |
|
| 587 |
test_imulw2(0x1234001d, 45); |
| 588 |
test_imulw2(23, -45); |
| 589 |
test_imulw2(0x8000, 0x8000); |
| 590 |
test_imulw2(0x100, 0x100); |
| 591 |
|
| 592 |
test_imull2(0x1234001d, 45); |
| 593 |
test_imull2(23, -45); |
| 594 |
test_imull2(0x80000000, 0x80000000); |
| 595 |
test_imull2(0x10000, 0x10000); |
| 596 |
|
| 597 |
TEST_IMUL_IM("w", "w", 45, 0x1234); |
| 598 |
TEST_IMUL_IM("w", "w", -45, 23); |
| 599 |
TEST_IMUL_IM("w", "w", 0x8000, 0x80000000); |
| 600 |
TEST_IMUL_IM("w", "w", 0x7fff, 0x1000); |
| 601 |
|
| 602 |
TEST_IMUL_IM("l", "k", 45, 0x1234); |
| 603 |
TEST_IMUL_IM("l", "k", -45, 23); |
| 604 |
TEST_IMUL_IM("l", "k", 0x8000, 0x80000000); |
| 605 |
TEST_IMUL_IM("l", "k", 0x7fff, 0x1000); |
| 606 |
|
| 607 |
test_idivb(0x12341678, 0x127e); |
| 608 |
test_idivb(0x43210123, -5); |
| 609 |
test_idivb(0x12340004, -1); |
| 610 |
|
| 611 |
test_idivw(0, 0x12345678, 12347); |
| 612 |
test_idivw(0, -23223, -45); |
| 613 |
test_idivw(0, 0x12348000, -1); |
| 614 |
test_idivw(0x12343, 0x12345678, 0x81238567); |
| 615 |
|
| 616 |
test_idivl(0, 0x12345678, 12347); |
| 617 |
test_idivl(0, -233223, -45); |
| 618 |
test_idivl(0, 0x80000000, -1); |
| 619 |
test_idivl(0x12343, 0x12345678, 0x81234567); |
| 620 |
|
| 621 |
test_divb(0x12341678, 0x127e); |
| 622 |
test_divb(0x43210123, -5); |
| 623 |
test_divb(0x12340004, -1); |
| 624 |
|
| 625 |
test_divw(0, 0x12345678, 12347); |
| 626 |
test_divw(0, -23223, -45); |
| 627 |
test_divw(0, 0x12348000, -1); |
| 628 |
test_divw(0x12343, 0x12345678, 0x81238567); |
| 629 |
|
| 630 |
test_divl(0, 0x12345678, 12347); |
| 631 |
test_divl(0, -233223, -45); |
| 632 |
test_divl(0, 0x80000000, -1); |
| 633 |
test_divl(0x12343, 0x12345678, 0x81234567); |
| 634 |
|
| 635 |
#if defined(__x86_64__)
|
| 636 |
test_imulq(0, 0x1234001d1234001d, 45); |
| 637 |
test_imulq(0, 23, -45); |
| 638 |
test_imulq(0, 0x8000000000000000, 0x8000000000000000); |
| 639 |
test_imulq(0, 0x100000000, 0x100000000); |
| 640 |
|
| 641 |
test_mulq(0, 0x1234001d1234001d, 45); |
| 642 |
test_mulq(0, 23, -45); |
| 643 |
test_mulq(0, 0x8000000000000000, 0x8000000000000000); |
| 644 |
test_mulq(0, 0x100000000, 0x100000000); |
| 645 |
|
| 646 |
test_imulq2(0x1234001d1234001d, 45); |
| 647 |
test_imulq2(23, -45); |
| 648 |
test_imulq2(0x8000000000000000, 0x8000000000000000); |
| 649 |
test_imulq2(0x100000000, 0x100000000); |
| 650 |
|
| 651 |
TEST_IMUL_IM("q", "", 45, 0x12341234); |
| 652 |
TEST_IMUL_IM("q", "", -45, 23); |
| 653 |
TEST_IMUL_IM("q", "", 0x8000, 0x8000000000000000); |
| 654 |
TEST_IMUL_IM("q", "", 0x7fff, 0x10000000); |
| 655 |
|
| 656 |
test_idivq(0, 0x12345678abcdef, 12347); |
| 657 |
test_idivq(0, -233223, -45); |
| 658 |
test_idivq(0, 0x8000000000000000, -1); |
| 659 |
test_idivq(0x12343, 0x12345678, 0x81234567); |
| 660 |
|
| 661 |
test_divq(0, 0x12345678abcdef, 12347); |
| 662 |
test_divq(0, -233223, -45); |
| 663 |
test_divq(0, 0x8000000000000000, -1); |
| 664 |
test_divq(0x12343, 0x12345678, 0x81234567); |
| 665 |
#endif
|
| 666 |
} |
| 667 |
|
| 668 |
#define TEST_BSX(op, size, op0)\
|
| 669 |
{\
|
| 670 |
long res, val, resz;\
|
| 671 |
val = op0;\ |
| 672 |
asm("xor %1, %1\n"\ |
| 673 |
"mov $0x12345678, %0\n"\
|
| 674 |
#op " %" size "2, %" size "0 ; setz %b1" \ |
| 675 |
: "=r" (res), "=q" (resz)\ |
| 676 |
: "g" (val));\
|
| 677 |
printf("%-10s A=" FMTLX " R=" FMTLX " %ld\n", #op, val, res, resz);\ |
| 678 |
} |
| 679 |
|
| 680 |
void test_bsx(void) |
| 681 |
{
|
| 682 |
TEST_BSX(bsrw, "w", 0); |
| 683 |
TEST_BSX(bsrw, "w", 0x12340128); |
| 684 |
TEST_BSX(bsfw, "w", 0); |
| 685 |
TEST_BSX(bsfw, "w", 0x12340128); |
| 686 |
TEST_BSX(bsrl, "k", 0); |
| 687 |
TEST_BSX(bsrl, "k", 0x00340128); |
| 688 |
TEST_BSX(bsfl, "k", 0); |
| 689 |
TEST_BSX(bsfl, "k", 0x00340128); |
| 690 |
#if defined(__x86_64__)
|
| 691 |
TEST_BSX(bsrq, "", 0); |
| 692 |
TEST_BSX(bsrq, "", 0x003401281234); |
| 693 |
TEST_BSX(bsfq, "", 0); |
| 694 |
TEST_BSX(bsfq, "", 0x003401281234); |
| 695 |
#endif
|
| 696 |
} |
| 697 |
|
| 698 |
/**********************************************/
|
| 699 |
|
| 700 |
void test_fops(double a, double b) |
| 701 |
{
|
| 702 |
printf("a=%f b=%f a+b=%f\n", a, b, a + b);
|
| 703 |
printf("a=%f b=%f a-b=%f\n", a, b, a - b);
|
| 704 |
printf("a=%f b=%f a*b=%f\n", a, b, a * b);
|
| 705 |
printf("a=%f b=%f a/b=%f\n", a, b, a / b);
|
| 706 |
printf("a=%f b=%f fmod(a, b)=%f\n", a, b, fmod(a, b));
|
| 707 |
printf("a=%f sqrt(a)=%f\n", a, sqrt(a));
|
| 708 |
printf("a=%f sin(a)=%f\n", a, sin(a));
|
| 709 |
printf("a=%f cos(a)=%f\n", a, cos(a));
|
| 710 |
printf("a=%f tan(a)=%f\n", a, tan(a));
|
| 711 |
printf("a=%f log(a)=%f\n", a, log(a));
|
| 712 |
printf("a=%f exp(a)=%f\n", a, exp(a));
|
| 713 |
printf("a=%f b=%f atan2(a, b)=%f\n", a, b, atan2(a, b));
|
| 714 |
/* just to test some op combining */
|
| 715 |
printf("a=%f asin(sin(a))=%f\n", a, asin(sin(a)));
|
| 716 |
printf("a=%f acos(cos(a))=%f\n", a, acos(cos(a)));
|
| 717 |
printf("a=%f atan(tan(a))=%f\n", a, atan(tan(a)));
|
| 718 |
|
| 719 |
} |
| 720 |
|
| 721 |
void test_fcmp(double a, double b) |
| 722 |
{
|
| 723 |
printf("(%f<%f)=%d\n",
|
| 724 |
a, b, a < b); |
| 725 |
printf("(%f<=%f)=%d\n",
|
| 726 |
a, b, a <= b); |
| 727 |
printf("(%f==%f)=%d\n",
|
| 728 |
a, b, a == b); |
| 729 |
printf("(%f>%f)=%d\n",
|
| 730 |
a, b, a > b); |
| 731 |
printf("(%f<=%f)=%d\n",
|
| 732 |
a, b, a >= b); |
| 733 |
if (TEST_FCOMI) {
|
| 734 |
long eflags;
|
| 735 |
/* test f(u)comi instruction */
|
| 736 |
asm("fcomi %2, %1\n" |
| 737 |
"pushf\n"
|
| 738 |
"pop %0\n"
|
| 739 |
: "=r" (eflags)
|
| 740 |
: "t" (a), "u" (b)); |
| 741 |
printf("fcomi(%f %f)=%08lx\n", a, b, eflags & (CC_Z | CC_P | CC_C));
|
| 742 |
} |
| 743 |
} |
| 744 |
|
| 745 |
void test_fcvt(double a) |
| 746 |
{
|
| 747 |
float fa;
|
| 748 |
long double la; |
| 749 |
int16_t fpuc; |
| 750 |
int i;
|
| 751 |
int64_t lla; |
| 752 |
int ia;
|
| 753 |
int16_t wa; |
| 754 |
double ra;
|
| 755 |
|
| 756 |
fa = a; |
| 757 |
la = a; |
| 758 |
printf("(float)%f = %f\n", a, fa);
|
| 759 |
printf("(long double)%f = %Lf\n", a, la);
|
| 760 |
printf("a=" FMT64X "\n", *(uint64_t *)&a); |
| 761 |
printf("la=" FMT64X " %04x\n", *(uint64_t *)&la, |
| 762 |
*(unsigned short *)((char *)(&la) + 8)); |
| 763 |
|
| 764 |
/* test all roundings */
|
| 765 |
asm volatile ("fstcw %0" : "=m" (fpuc)); |
| 766 |
for(i=0;i<4;i++) { |
| 767 |
asm volatile ("fldcw %0" : : "m" ((fpuc & ~0x0c00) | (i << 10))); |
| 768 |
asm volatile ("fist %0" : "=m" (wa) : "t" (a)); |
| 769 |
asm volatile ("fistl %0" : "=m" (ia) : "t" (a)); |
| 770 |
asm volatile ("fistpll %0" : "=m" (lla) : "t" (a) : "st"); |
| 771 |
asm volatile ("frndint ; fstl %0" : "=m" (ra) : "t" (a)); |
| 772 |
asm volatile ("fldcw %0" : : "m" (fpuc)); |
| 773 |
printf("(short)a = %d\n", wa);
|
| 774 |
printf("(int)a = %d\n", ia);
|
| 775 |
printf("(int64_t)a = " FMT64X "\n", lla); |
| 776 |
printf("rint(a) = %f\n", ra);
|
| 777 |
} |
| 778 |
} |
| 779 |
|
| 780 |
#define TEST(N) \
|
| 781 |
asm("fld" #N : "=t" (a)); \ |
| 782 |
printf("fld" #N "= %f\n", a); |
| 783 |
|
| 784 |
void test_fconst(void) |
| 785 |
{
|
| 786 |
double a;
|
| 787 |
TEST(1);
|
| 788 |
TEST(l2t); |
| 789 |
TEST(l2e); |
| 790 |
TEST(pi); |
| 791 |
TEST(lg2); |
| 792 |
TEST(ln2); |
| 793 |
TEST(z); |
| 794 |
} |
| 795 |
|
| 796 |
void test_fbcd(double a) |
| 797 |
{
|
| 798 |
unsigned short bcd[5]; |
| 799 |
double b;
|
| 800 |
|
| 801 |
asm("fbstp %0" : "=m" (bcd[0]) : "t" (a) : "st"); |
| 802 |
asm("fbld %1" : "=t" (b) : "m" (bcd[0])); |
| 803 |
printf("a=%f bcd=%04x%04x%04x%04x%04x b=%f\n",
|
| 804 |
a, bcd[4], bcd[3], bcd[2], bcd[1], bcd[0], b); |
| 805 |
} |
| 806 |
|
| 807 |
#define TEST_ENV(env, save, restore)\
|
| 808 |
{\
|
| 809 |
memset((env), 0xaa, sizeof(*(env)));\ |
| 810 |
for(i=0;i<5;i++)\ |
| 811 |
asm volatile ("fldl %0" : : "m" (dtab[i]));\ |
| 812 |
asm volatile (save " %0\n" : : "m" (*(env)));\ |
| 813 |
asm volatile (restore " %0\n": : "m" (*(env)));\ |
| 814 |
for(i=0;i<5;i++)\ |
| 815 |
asm volatile ("fstpl %0" : "=m" (rtab[i]));\ |
| 816 |
for(i=0;i<5;i++)\ |
| 817 |
printf("res[%d]=%f\n", i, rtab[i]);\
|
| 818 |
printf("fpuc=%04x fpus=%04x fptag=%04x\n",\
|
| 819 |
(env)->fpuc,\ |
| 820 |
(env)->fpus & 0xff00,\
|
| 821 |
(env)->fptag);\ |
| 822 |
} |
| 823 |
|
| 824 |
void test_fenv(void) |
| 825 |
{
|
| 826 |
struct __attribute__((packed)) {
|
| 827 |
uint16_t fpuc; |
| 828 |
uint16_t dummy1; |
| 829 |
uint16_t fpus; |
| 830 |
uint16_t dummy2; |
| 831 |
uint16_t fptag; |
| 832 |
uint16_t dummy3; |
| 833 |
uint32_t ignored[4];
|
| 834 |
long double fpregs[8]; |
| 835 |
} float_env32; |
| 836 |
struct __attribute__((packed)) {
|
| 837 |
uint16_t fpuc; |
| 838 |
uint16_t fpus; |
| 839 |
uint16_t fptag; |
| 840 |
uint16_t ignored[4];
|
| 841 |
long double fpregs[8]; |
| 842 |
} float_env16; |
| 843 |
double dtab[8]; |
| 844 |
double rtab[8]; |
| 845 |
int i;
|
| 846 |
|
| 847 |
for(i=0;i<8;i++) |
| 848 |
dtab[i] = i + 1;
|
| 849 |
|
| 850 |
TEST_ENV(&float_env16, "data16 fnstenv", "data16 fldenv"); |
| 851 |
TEST_ENV(&float_env16, "data16 fnsave", "data16 frstor"); |
| 852 |
TEST_ENV(&float_env32, "fnstenv", "fldenv"); |
| 853 |
TEST_ENV(&float_env32, "fnsave", "frstor"); |
| 854 |
|
| 855 |
/* test for ffree */
|
| 856 |
for(i=0;i<5;i++) |
| 857 |
asm volatile ("fldl %0" : : "m" (dtab[i])); |
| 858 |
asm volatile("ffree %st(2)"); |
| 859 |
asm volatile ("fnstenv %0\n" : : "m" (float_env32)); |
| 860 |
asm volatile ("fninit"); |
| 861 |
printf("fptag=%04x\n", float_env32.fptag);
|
| 862 |
} |
| 863 |
|
| 864 |
|
| 865 |
#define TEST_FCMOV(a, b, eflags, CC)\
|
| 866 |
{\
|
| 867 |
double res;\
|
| 868 |
asm("push %3\n"\ |
| 869 |
"popf\n"\
|
| 870 |
"fcmov" CC " %2, %0\n"\ |
| 871 |
: "=t" (res)\
|
| 872 |
: "0" (a), "u" (b), "g" (eflags));\ |
| 873 |
printf("fcmov%s eflags=0x%04lx-> %f\n", \
|
| 874 |
CC, (long)eflags, res);\
|
| 875 |
} |
| 876 |
|
| 877 |
void test_fcmov(void) |
| 878 |
{
|
| 879 |
double a, b;
|
| 880 |
long eflags, i;
|
| 881 |
|
| 882 |
a = 1.0; |
| 883 |
b = 2.0; |
| 884 |
for(i = 0; i < 4; i++) { |
| 885 |
eflags = 0;
|
| 886 |
if (i & 1) |
| 887 |
eflags |= CC_C; |
| 888 |
if (i & 2) |
| 889 |
eflags |= CC_Z; |
| 890 |
TEST_FCMOV(a, b, eflags, "b");
|
| 891 |
TEST_FCMOV(a, b, eflags, "e");
|
| 892 |
TEST_FCMOV(a, b, eflags, "be");
|
| 893 |
TEST_FCMOV(a, b, eflags, "nb");
|
| 894 |
TEST_FCMOV(a, b, eflags, "ne");
|
| 895 |
TEST_FCMOV(a, b, eflags, "nbe");
|
| 896 |
} |
| 897 |
TEST_FCMOV(a, b, 0, "u"); |
| 898 |
TEST_FCMOV(a, b, CC_P, "u");
|
| 899 |
TEST_FCMOV(a, b, 0, "nu"); |
| 900 |
TEST_FCMOV(a, b, CC_P, "nu");
|
| 901 |
} |
| 902 |
|
| 903 |
void test_floats(void) |
| 904 |
{
|
| 905 |
test_fops(2, 3); |
| 906 |
test_fops(1.4, -5); |
| 907 |
test_fcmp(2, -1); |
| 908 |
test_fcmp(2, 2); |
| 909 |
test_fcmp(2, 3); |
| 910 |
test_fcvt(0.5); |
| 911 |
test_fcvt(-0.5); |
| 912 |
test_fcvt(1.0/7.0); |
| 913 |
test_fcvt(-1.0/9.0); |
| 914 |
test_fcvt(32768);
|
| 915 |
test_fcvt(-1e20);
|
| 916 |
test_fconst(); |
| 917 |
test_fbcd(1234567890123456);
|
| 918 |
test_fbcd(-123451234567890);
|
| 919 |
test_fenv(); |
| 920 |
if (TEST_CMOV) {
|
| 921 |
test_fcmov(); |
| 922 |
} |
| 923 |
} |
| 924 |
|
| 925 |
/**********************************************/
|
| 926 |
#if !defined(__x86_64__)
|
| 927 |
|
| 928 |
#define TEST_BCD(op, op0, cc_in, cc_mask)\
|
| 929 |
{\
|
| 930 |
int res, flags;\
|
| 931 |
res = op0;\ |
| 932 |
flags = cc_in;\ |
| 933 |
asm ("push %3\n\t"\ |
| 934 |
"popf\n\t"\
|
| 935 |
#op "\n\t"\ |
| 936 |
"pushf\n\t"\
|
| 937 |
"pop %1\n\t"\
|
| 938 |
: "=a" (res), "=g" (flags)\ |
| 939 |
: "0" (res), "1" (flags));\ |
| 940 |
printf("%-10s A=%08x R=%08x CCIN=%04x CC=%04x\n",\
|
| 941 |
#op, op0, res, cc_in, flags & cc_mask);\
|
| 942 |
} |
| 943 |
|
| 944 |
void test_bcd(void) |
| 945 |
{
|
| 946 |
TEST_BCD(daa, 0x12340503, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 947 |
TEST_BCD(daa, 0x12340506, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 948 |
TEST_BCD(daa, 0x12340507, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 949 |
TEST_BCD(daa, 0x12340559, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 950 |
TEST_BCD(daa, 0x12340560, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 951 |
TEST_BCD(daa, 0x1234059f, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 952 |
TEST_BCD(daa, 0x123405a0, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 953 |
TEST_BCD(daa, 0x12340503, 0, (CC_C | CC_P | CC_Z | CC_S | CC_A)); |
| 954 |
TEST_BCD(daa, 0x12340506, 0, (CC_C | CC_P | CC_Z | CC_S | CC_A)); |
| 955 |
TEST_BCD(daa, 0x12340503, CC_C, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 956 |
TEST_BCD(daa, 0x12340506, CC_C, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 957 |
TEST_BCD(daa, 0x12340503, CC_C | CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 958 |
TEST_BCD(daa, 0x12340506, CC_C | CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 959 |
|
| 960 |
TEST_BCD(das, 0x12340503, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 961 |
TEST_BCD(das, 0x12340506, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 962 |
TEST_BCD(das, 0x12340507, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 963 |
TEST_BCD(das, 0x12340559, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 964 |
TEST_BCD(das, 0x12340560, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 965 |
TEST_BCD(das, 0x1234059f, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 966 |
TEST_BCD(das, 0x123405a0, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 967 |
TEST_BCD(das, 0x12340503, 0, (CC_C | CC_P | CC_Z | CC_S | CC_A)); |
| 968 |
TEST_BCD(das, 0x12340506, 0, (CC_C | CC_P | CC_Z | CC_S | CC_A)); |
| 969 |
TEST_BCD(das, 0x12340503, CC_C, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 970 |
TEST_BCD(das, 0x12340506, CC_C, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 971 |
TEST_BCD(das, 0x12340503, CC_C | CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 972 |
TEST_BCD(das, 0x12340506, CC_C | CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_A));
|
| 973 |
|
| 974 |
TEST_BCD(aaa, 0x12340205, CC_A, (CC_C | CC_A));
|
| 975 |
TEST_BCD(aaa, 0x12340306, CC_A, (CC_C | CC_A));
|
| 976 |
TEST_BCD(aaa, 0x1234040a, CC_A, (CC_C | CC_A));
|
| 977 |
TEST_BCD(aaa, 0x123405fa, CC_A, (CC_C | CC_A));
|
| 978 |
TEST_BCD(aaa, 0x12340205, 0, (CC_C | CC_A)); |
| 979 |
TEST_BCD(aaa, 0x12340306, 0, (CC_C | CC_A)); |
| 980 |
TEST_BCD(aaa, 0x1234040a, 0, (CC_C | CC_A)); |
| 981 |
TEST_BCD(aaa, 0x123405fa, 0, (CC_C | CC_A)); |
| 982 |
|
| 983 |
TEST_BCD(aas, 0x12340205, CC_A, (CC_C | CC_A));
|
| 984 |
TEST_BCD(aas, 0x12340306, CC_A, (CC_C | CC_A));
|
| 985 |
TEST_BCD(aas, 0x1234040a, CC_A, (CC_C | CC_A));
|
| 986 |
TEST_BCD(aas, 0x123405fa, CC_A, (CC_C | CC_A));
|
| 987 |
TEST_BCD(aas, 0x12340205, 0, (CC_C | CC_A)); |
| 988 |
TEST_BCD(aas, 0x12340306, 0, (CC_C | CC_A)); |
| 989 |
TEST_BCD(aas, 0x1234040a, 0, (CC_C | CC_A)); |
| 990 |
TEST_BCD(aas, 0x123405fa, 0, (CC_C | CC_A)); |
| 991 |
|
| 992 |
TEST_BCD(aam, 0x12340547, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A));
|
| 993 |
TEST_BCD(aad, 0x12340407, CC_A, (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A));
|
| 994 |
} |
| 995 |
#endif
|
| 996 |
|
| 997 |
#define TEST_XCHG(op, size, opconst)\
|
| 998 |
{\
|
| 999 |
long op0, op1;\
|
| 1000 |
op0 = i2l(0x12345678);\
|
| 1001 |
op1 = i2l(0xfbca7654);\
|
| 1002 |
asm(#op " %" size "0, %" size "1" \ |
| 1003 |
: "=q" (op0), opconst (op1) \
|
| 1004 |
: "0" (op0), "1" (op1));\ |
| 1005 |
printf("%-10s A=" FMTLX " B=" FMTLX "\n",\ |
| 1006 |
#op, op0, op1);\
|
| 1007 |
} |
| 1008 |
|
| 1009 |
#define TEST_CMPXCHG(op, size, opconst, eax)\
|
| 1010 |
{\
|
| 1011 |
long op0, op1, op2;\
|
| 1012 |
op0 = i2l(0x12345678);\
|
| 1013 |
op1 = i2l(0xfbca7654);\
|
| 1014 |
op2 = i2l(eax);\ |
| 1015 |
asm(#op " %" size "0, %" size "1" \ |
| 1016 |
: "=q" (op0), opconst (op1) \
|
| 1017 |
: "0" (op0), "1" (op1), "a" (op2));\ |
| 1018 |
printf("%-10s EAX=" FMTLX " A=" FMTLX " C=" FMTLX "\n",\ |
| 1019 |
#op, op2, op0, op1);\
|
| 1020 |
} |
| 1021 |
|
| 1022 |
void test_xchg(void) |
| 1023 |
{
|
| 1024 |
#if defined(__x86_64__)
|
| 1025 |
TEST_XCHG(xchgq, "", "=q"); |
| 1026 |
#endif
|
| 1027 |
TEST_XCHG(xchgl, "k", "=q"); |
| 1028 |
TEST_XCHG(xchgw, "w", "=q"); |
| 1029 |
TEST_XCHG(xchgb, "b", "=q"); |
| 1030 |
|
| 1031 |
#if defined(__x86_64__)
|
| 1032 |
TEST_XCHG(xchgq, "", "=m"); |
| 1033 |
#endif
|
| 1034 |
TEST_XCHG(xchgl, "k", "=m"); |
| 1035 |
TEST_XCHG(xchgw, "w", "=m"); |
| 1036 |
TEST_XCHG(xchgb, "b", "=m"); |
| 1037 |
|
| 1038 |
#if defined(__x86_64__)
|
| 1039 |
TEST_XCHG(xaddq, "", "=q"); |
| 1040 |
#endif
|
| 1041 |
TEST_XCHG(xaddl, "k", "=q"); |
| 1042 |
TEST_XCHG(xaddw, "w", "=q"); |
| 1043 |
TEST_XCHG(xaddb, "b", "=q"); |
| 1044 |
|
| 1045 |
{
|
| 1046 |
int res;
|
| 1047 |
res = 0x12345678;
|
| 1048 |
asm("xaddl %1, %0" : "=r" (res) : "0" (res)); |
| 1049 |
printf("xaddl same res=%08x\n", res);
|
| 1050 |
} |
| 1051 |
|
| 1052 |
#if defined(__x86_64__)
|
| 1053 |
TEST_XCHG(xaddq, "", "=m"); |
| 1054 |
#endif
|
| 1055 |
TEST_XCHG(xaddl, "k", "=m"); |
| 1056 |
TEST_XCHG(xaddw, "w", "=m"); |
| 1057 |
TEST_XCHG(xaddb, "b", "=m"); |
| 1058 |
|
| 1059 |
#if defined(__x86_64__)
|
| 1060 |
TEST_CMPXCHG(cmpxchgq, "", "=q", 0xfbca7654); |
| 1061 |
#endif
|
| 1062 |
TEST_CMPXCHG(cmpxchgl, "k", "=q", 0xfbca7654); |
| 1063 |
TEST_CMPXCHG(cmpxchgw, "w", "=q", 0xfbca7654); |
| 1064 |
TEST_CMPXCHG(cmpxchgb, "b", "=q", 0xfbca7654); |
| 1065 |
|
| 1066 |
#if defined(__x86_64__)
|
| 1067 |
TEST_CMPXCHG(cmpxchgq, "", "=q", 0xfffefdfc); |
| 1068 |
#endif
|
| 1069 |
TEST_CMPXCHG(cmpxchgl, "k", "=q", 0xfffefdfc); |
| 1070 |
TEST_CMPXCHG(cmpxchgw, "w", "=q", 0xfffefdfc); |
| 1071 |
TEST_CMPXCHG(cmpxchgb, "b", "=q", 0xfffefdfc); |
| 1072 |
|
| 1073 |
#if defined(__x86_64__)
|
| 1074 |
TEST_CMPXCHG(cmpxchgq, "", "=m", 0xfbca7654); |
| 1075 |
#endif
|
| 1076 |
TEST_CMPXCHG(cmpxchgl, "k", "=m", 0xfbca7654); |
| 1077 |
TEST_CMPXCHG(cmpxchgw, "w", "=m", 0xfbca7654); |
| 1078 |
TEST_CMPXCHG(cmpxchgb, "b", "=m", 0xfbca7654); |
| 1079 |
|
| 1080 |
#if defined(__x86_64__)
|
| 1081 |
TEST_CMPXCHG(cmpxchgq, "", "=m", 0xfffefdfc); |
| 1082 |
#endif
|
| 1083 |
TEST_CMPXCHG(cmpxchgl, "k", "=m", 0xfffefdfc); |
| 1084 |
TEST_CMPXCHG(cmpxchgw, "w", "=m", 0xfffefdfc); |
| 1085 |
TEST_CMPXCHG(cmpxchgb, "b", "=m", 0xfffefdfc); |
| 1086 |
|
| 1087 |
{
|
| 1088 |
uint64_t op0, op1, op2; |
| 1089 |
long i, eflags;
|
| 1090 |
|
| 1091 |
for(i = 0; i < 2; i++) { |
| 1092 |
op0 = 0x123456789abcd;
|
| 1093 |
if (i == 0) |
| 1094 |
op1 = 0xfbca765423456;
|
| 1095 |
else
|
| 1096 |
op1 = op0; |
| 1097 |
op2 = 0x6532432432434;
|
| 1098 |
asm("cmpxchg8b %1\n" |
| 1099 |
"pushf\n"
|
| 1100 |
"pop %2\n"
|
| 1101 |
: "=A" (op0), "=m" (op1), "=g" (eflags) |
| 1102 |
: "0" (op0), "m" (op1), "b" ((int)op2), "c" ((int)(op2 >> 32))); |
| 1103 |
printf("cmpxchg8b: op0=" FMT64X " op1=" FMT64X " CC=%02lx\n", |
| 1104 |
op0, op1, eflags & CC_Z); |
| 1105 |
} |
| 1106 |
} |
| 1107 |
} |
| 1108 |
|
| 1109 |
#ifdef TEST_SEGS
|
| 1110 |
/**********************************************/
|
| 1111 |
/* segmentation tests */
|
| 1112 |
|
| 1113 |
#include <asm/ldt.h> |
| 1114 |
#include <linux/unistd.h> |
| 1115 |
#include <linux/version.h> |
| 1116 |
|
| 1117 |
_syscall3(int, modify_ldt, int, func, void *, ptr, unsigned long, bytecount) |
| 1118 |
|
| 1119 |
#if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 5, 66) |
| 1120 |
#define modify_ldt_ldt_s user_desc
|
| 1121 |
#endif
|
| 1122 |
|
| 1123 |
#define MK_SEL(n) (((n) << 3) | 7) |
| 1124 |
|
| 1125 |
uint8_t seg_data1[4096];
|
| 1126 |
uint8_t seg_data2[4096];
|
| 1127 |
|
| 1128 |
#define TEST_LR(op, size, seg, mask)\
|
| 1129 |
{\
|
| 1130 |
int res, res2;\
|
| 1131 |
res = 0x12345678;\
|
| 1132 |
asm (op " %" size "2, %" size "0\n" \ |
| 1133 |
"movl $0, %1\n"\
|
| 1134 |
"jnz 1f\n"\
|
| 1135 |
"movl $1, %1\n"\
|
| 1136 |
"1:\n"\
|
| 1137 |
: "=r" (res), "=r" (res2) : "m" (seg), "0" (res));\ |
| 1138 |
printf(op ": Z=%d %08x\n", res2, res & ~(mask));\
|
| 1139 |
} |
| 1140 |
|
| 1141 |
/* NOTE: we use Linux modify_ldt syscall */
|
| 1142 |
void test_segs(void) |
| 1143 |
{
|
| 1144 |
struct modify_ldt_ldt_s ldt;
|
| 1145 |
long long ldt_table[3]; |
| 1146 |
int res, res2;
|
| 1147 |
char tmp;
|
| 1148 |
struct {
|
| 1149 |
uint32_t offset; |
| 1150 |
uint16_t seg; |
| 1151 |
} __attribute__((packed)) segoff; |
| 1152 |
|
| 1153 |
ldt.entry_number = 1;
|
| 1154 |
ldt.base_addr = (unsigned long)&seg_data1; |
| 1155 |
ldt.limit = (sizeof(seg_data1) + 0xfff) >> 12; |
| 1156 |
ldt.seg_32bit = 1;
|
| 1157 |
ldt.contents = MODIFY_LDT_CONTENTS_DATA; |
| 1158 |
ldt.read_exec_only = 0;
|
| 1159 |
ldt.limit_in_pages = 1;
|
| 1160 |
ldt.seg_not_present = 0;
|
| 1161 |
ldt.useable = 1;
|
| 1162 |
modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */ |
| 1163 |
|
| 1164 |
ldt.entry_number = 2;
|
| 1165 |
ldt.base_addr = (unsigned long)&seg_data2; |
| 1166 |
ldt.limit = (sizeof(seg_data2) + 0xfff) >> 12; |
| 1167 |
ldt.seg_32bit = 1;
|
| 1168 |
ldt.contents = MODIFY_LDT_CONTENTS_DATA; |
| 1169 |
ldt.read_exec_only = 0;
|
| 1170 |
ldt.limit_in_pages = 1;
|
| 1171 |
ldt.seg_not_present = 0;
|
| 1172 |
ldt.useable = 1;
|
| 1173 |
modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */ |
| 1174 |
|
| 1175 |
modify_ldt(0, &ldt_table, sizeof(ldt_table)); /* read ldt entries */ |
| 1176 |
#if 0
|
| 1177 |
{
|
| 1178 |
int i;
|
| 1179 |
for(i=0;i<3;i++)
|
| 1180 |
printf("%d: %016Lx\n", i, ldt_table[i]);
|
| 1181 |
}
|
| 1182 |
#endif
|
| 1183 |
/* do some tests with fs or gs */
|
| 1184 |
asm volatile ("movl %0, %%fs" : : "r" (MK_SEL(1))); |
| 1185 |
|
| 1186 |
seg_data1[1] = 0xaa; |
| 1187 |
seg_data2[1] = 0x55; |
| 1188 |
|
| 1189 |
asm volatile ("fs movzbl 0x1, %0" : "=r" (res)); |
| 1190 |
printf("FS[1] = %02x\n", res);
|
| 1191 |
|
| 1192 |
asm volatile ("pushl %%gs\n" |
| 1193 |
"movl %1, %%gs\n"
|
| 1194 |
"gs movzbl 0x1, %0\n"
|
| 1195 |
"popl %%gs\n"
|
| 1196 |
: "=r" (res)
|
| 1197 |
: "r" (MK_SEL(2))); |
| 1198 |
printf("GS[1] = %02x\n", res);
|
| 1199 |
|
| 1200 |
/* tests with ds/ss (implicit segment case) */
|
| 1201 |
tmp = 0xa5;
|
| 1202 |
asm volatile ("pushl %%ebp\n\t" |
| 1203 |
"pushl %%ds\n\t"
|
| 1204 |
"movl %2, %%ds\n\t"
|
| 1205 |
"movl %3, %%ebp\n\t"
|
| 1206 |
"movzbl 0x1, %0\n\t"
|
| 1207 |
"movzbl (%%ebp), %1\n\t"
|
| 1208 |
"popl %%ds\n\t"
|
| 1209 |
"popl %%ebp\n\t"
|
| 1210 |
: "=r" (res), "=r" (res2) |
| 1211 |
: "r" (MK_SEL(1)), "r" (&tmp)); |
| 1212 |
printf("DS[1] = %02x\n", res);
|
| 1213 |
printf("SS[tmp] = %02x\n", res2);
|
| 1214 |
|
| 1215 |
segoff.seg = MK_SEL(2);
|
| 1216 |
segoff.offset = 0xabcdef12;
|
| 1217 |
asm volatile("lfs %2, %0\n\t" |
| 1218 |
"movl %%fs, %1\n\t"
|
| 1219 |
: "=r" (res), "=g" (res2) |
| 1220 |
: "m" (segoff));
|
| 1221 |
printf("FS:reg = %04x:%08x\n", res2, res);
|
| 1222 |
|
| 1223 |
TEST_LR("larw", "w", MK_SEL(2), 0x0100); |
| 1224 |
TEST_LR("larl", "", MK_SEL(2), 0x0100); |
| 1225 |
TEST_LR("lslw", "w", MK_SEL(2), 0); |
| 1226 |
TEST_LR("lsll", "", MK_SEL(2), 0); |
| 1227 |
|
| 1228 |
TEST_LR("larw", "w", 0xfff8, 0); |
| 1229 |
TEST_LR("larl", "", 0xfff8, 0); |
| 1230 |
TEST_LR("lslw", "w", 0xfff8, 0); |
| 1231 |
TEST_LR("lsll", "", 0xfff8, 0); |
| 1232 |
} |
| 1233 |
|
| 1234 |
/* 16 bit code test */
|
| 1235 |
extern char code16_start, code16_end; |
| 1236 |
extern char code16_func1; |
| 1237 |
extern char code16_func2; |
| 1238 |
extern char code16_func3; |
| 1239 |
|
| 1240 |
void test_code16(void) |
| 1241 |
{
|
| 1242 |
struct modify_ldt_ldt_s ldt;
|
| 1243 |
int res, res2;
|
| 1244 |
|
| 1245 |
/* build a code segment */
|
| 1246 |
ldt.entry_number = 1;
|
| 1247 |
ldt.base_addr = (unsigned long)&code16_start; |
| 1248 |
ldt.limit = &code16_end - &code16_start; |
| 1249 |
ldt.seg_32bit = 0;
|
| 1250 |
ldt.contents = MODIFY_LDT_CONTENTS_CODE; |
| 1251 |
ldt.read_exec_only = 0;
|
| 1252 |
ldt.limit_in_pages = 0;
|
| 1253 |
ldt.seg_not_present = 0;
|
| 1254 |
ldt.useable = 1;
|
| 1255 |
modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */ |
| 1256 |
|
| 1257 |
/* call the first function */
|
| 1258 |
asm volatile ("lcall %1, %2" |
| 1259 |
: "=a" (res)
|
| 1260 |
: "i" (MK_SEL(1)), "i" (&code16_func1): "memory", "cc"); |
| 1261 |
printf("func1() = 0x%08x\n", res);
|
| 1262 |
asm volatile ("lcall %2, %3" |
| 1263 |
: "=a" (res), "=c" (res2) |
| 1264 |
: "i" (MK_SEL(1)), "i" (&code16_func2): "memory", "cc"); |
| 1265 |
printf("func2() = 0x%08x spdec=%d\n", res, res2);
|
| 1266 |
asm volatile ("lcall %1, %2" |
| 1267 |
: "=a" (res)
|
| 1268 |
: "i" (MK_SEL(1)), "i" (&code16_func3): "memory", "cc"); |
| 1269 |
printf("func3() = 0x%08x\n", res);
|
| 1270 |
} |
| 1271 |
#endif
|
| 1272 |
|
| 1273 |
extern char func_lret32; |
| 1274 |
extern char func_iret32; |
| 1275 |
|
| 1276 |
void test_misc(void) |
| 1277 |
{
|
| 1278 |
char table[256]; |
| 1279 |
long res, i;
|
| 1280 |
|
| 1281 |
for(i=0;i<256;i++) table[i] = 256 - i; |
| 1282 |
res = 0x12345678;
|
| 1283 |
asm ("xlat" : "=a" (res) : "b" (table), "0" (res)); |
| 1284 |
printf("xlat: EAX=" FMTLX "\n", res); |
| 1285 |
|
| 1286 |
#if !defined(__x86_64__)
|
| 1287 |
asm volatile ("push %%cs ; call %1" |
| 1288 |
: "=a" (res)
|
| 1289 |
: "m" (func_lret32): "memory", "cc"); |
| 1290 |
printf("func_lret32=" FMTLX "\n", res); |
| 1291 |
|
| 1292 |
asm volatile ("pushf ; push %%cs ; call %1" |
| 1293 |
: "=a" (res)
|
| 1294 |
: "m" (func_iret32): "memory", "cc"); |
| 1295 |
printf("func_iret32=" FMTLX "\n", res); |
| 1296 |
#endif
|
| 1297 |
|
| 1298 |
#if defined(__x86_64__)
|
| 1299 |
/* specific popl test */
|
| 1300 |
asm volatile ("push $12345432 ; push $0x9abcdef ; pop (%%rsp) ; pop %0" |
| 1301 |
: "=g" (res));
|
| 1302 |
printf("popl esp=" FMTLX "\n", res); |
| 1303 |
#else
|
| 1304 |
/* specific popl test */
|
| 1305 |
asm volatile ("pushl $12345432 ; pushl $0x9abcdef ; popl (%%esp) ; popl %0" |
| 1306 |
: "=g" (res));
|
| 1307 |
printf("popl esp=" FMTLX "\n", res); |
| 1308 |
|
| 1309 |
/* specific popw test */
|
| 1310 |
asm volatile ("pushl $12345432 ; pushl $0x9abcdef ; popw (%%esp) ; addl $2, %%esp ; popl %0" |
| 1311 |
: "=g" (res));
|
| 1312 |
printf("popw esp=" FMTLX "\n", res); |
| 1313 |
#endif
|
| 1314 |
} |
| 1315 |
|
| 1316 |
uint8_t str_buffer[4096];
|
| 1317 |
|
| 1318 |
#define TEST_STRING1(OP, size, DF, REP)\
|
| 1319 |
{\
|
| 1320 |
long esi, edi, eax, ecx, eflags;\
|
| 1321 |
\ |
| 1322 |
esi = (long)(str_buffer + sizeof(str_buffer) / 2);\ |
| 1323 |
edi = (long)(str_buffer + sizeof(str_buffer) / 2) + 16;\ |
| 1324 |
eax = i2l(0x12345678);\
|
| 1325 |
ecx = 17;\
|
| 1326 |
\ |
| 1327 |
asm volatile ("push $0\n\t"\ |
| 1328 |
"popf\n\t"\
|
| 1329 |
DF "\n\t"\
|
| 1330 |
REP #OP size "\n\t"\ |
| 1331 |
"cld\n\t"\
|
| 1332 |
"pushf\n\t"\
|
| 1333 |
"pop %4\n\t"\
|
| 1334 |
: "=S" (esi), "=D" (edi), "=a" (eax), "=c" (ecx), "=g" (eflags)\ |
| 1335 |
: "0" (esi), "1" (edi), "2" (eax), "3" (ecx));\ |
| 1336 |
printf("%-10s ESI=" FMTLX " EDI=" FMTLX " EAX=" FMTLX " ECX=" FMTLX " EFL=%04x\n",\ |
| 1337 |
REP #OP size, esi, edi, eax, ecx,\
|
| 1338 |
(int)(eflags & (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A)));\
|
| 1339 |
} |
| 1340 |
|
| 1341 |
#define TEST_STRING(OP, REP)\
|
| 1342 |
TEST_STRING1(OP, "b", "", REP);\ |
| 1343 |
TEST_STRING1(OP, "w", "", REP);\ |
| 1344 |
TEST_STRING1(OP, "l", "", REP);\ |
| 1345 |
X86_64_ONLY(TEST_STRING1(OP, "q", "", REP));\ |
| 1346 |
TEST_STRING1(OP, "b", "std", REP);\ |
| 1347 |
TEST_STRING1(OP, "w", "std", REP);\ |
| 1348 |
TEST_STRING1(OP, "l", "std", REP);\ |
| 1349 |
X86_64_ONLY(TEST_STRING1(OP, "q", "std", REP)) |
| 1350 |
|
| 1351 |
void test_string(void) |
| 1352 |
{
|
| 1353 |
int i;
|
| 1354 |
for(i = 0;i < sizeof(str_buffer); i++) |
| 1355 |
str_buffer[i] = i + 0x56;
|
| 1356 |
TEST_STRING(stos, "");
|
| 1357 |
TEST_STRING(stos, "rep ");
|
| 1358 |
TEST_STRING(lods, ""); /* to verify stos */ |
| 1359 |
TEST_STRING(lods, "rep ");
|
| 1360 |
TEST_STRING(movs, "");
|
| 1361 |
TEST_STRING(movs, "rep ");
|
| 1362 |
TEST_STRING(lods, ""); /* to verify stos */ |
| 1363 |
|
| 1364 |
/* XXX: better tests */
|
| 1365 |
TEST_STRING(scas, "");
|
| 1366 |
TEST_STRING(scas, "repz ");
|
| 1367 |
TEST_STRING(scas, "repnz ");
|
| 1368 |
TEST_STRING(cmps, "");
|
| 1369 |
TEST_STRING(cmps, "repz ");
|
| 1370 |
TEST_STRING(cmps, "repnz ");
|
| 1371 |
} |
| 1372 |
|
| 1373 |
#ifdef TEST_VM86
|
| 1374 |
/* VM86 test */
|
| 1375 |
|
| 1376 |
static inline void set_bit(uint8_t *a, unsigned int bit) |
| 1377 |
{
|
| 1378 |
a[bit / 8] |= (1 << (bit % 8)); |
| 1379 |
} |
| 1380 |
|
| 1381 |
static inline uint8_t *seg_to_linear(unsigned int seg, unsigned int reg) |
| 1382 |
{
|
| 1383 |
return (uint8_t *)((seg << 4) + (reg & 0xffff)); |
| 1384 |
} |
| 1385 |
|
| 1386 |
static inline void pushw(struct vm86_regs *r, int val) |
| 1387 |
{
|
| 1388 |
r->esp = (r->esp & ~0xffff) | ((r->esp - 2) & 0xffff); |
| 1389 |
*(uint16_t *)seg_to_linear(r->ss, r->esp) = val; |
| 1390 |
} |
| 1391 |
|
| 1392 |
#undef __syscall_return
|
| 1393 |
#define __syscall_return(type, res) \
|
| 1394 |
do { \
|
| 1395 |
return (type) (res); \
|
| 1396 |
} while (0) |
| 1397 |
|
| 1398 |
_syscall2(int, vm86, int, func, struct vm86plus_struct *, v86) |
| 1399 |
|
| 1400 |
extern char vm86_code_start; |
| 1401 |
extern char vm86_code_end; |
| 1402 |
|
| 1403 |
#define VM86_CODE_CS 0x100 |
| 1404 |
#define VM86_CODE_IP 0x100 |
| 1405 |
|
| 1406 |
void test_vm86(void) |
| 1407 |
{
|
| 1408 |
struct vm86plus_struct ctx;
|
| 1409 |
struct vm86_regs *r;
|
| 1410 |
uint8_t *vm86_mem; |
| 1411 |
int seg, ret;
|
| 1412 |
|
| 1413 |
vm86_mem = mmap((void *)0x00000000, 0x110000, |
| 1414 |
PROT_WRITE | PROT_READ | PROT_EXEC, |
| 1415 |
MAP_FIXED | MAP_ANON | MAP_PRIVATE, -1, 0); |
| 1416 |
if (vm86_mem == MAP_FAILED) {
|
| 1417 |
printf("ERROR: could not map vm86 memory");
|
| 1418 |
return;
|
| 1419 |
} |
| 1420 |
memset(&ctx, 0, sizeof(ctx)); |
| 1421 |
|
| 1422 |
/* init basic registers */
|
| 1423 |
r = &ctx.regs; |
| 1424 |
r->eip = VM86_CODE_IP; |
| 1425 |
r->esp = 0xfffe;
|
| 1426 |
seg = VM86_CODE_CS; |
| 1427 |
r->cs = seg; |
| 1428 |
r->ss = seg; |
| 1429 |
r->ds = seg; |
| 1430 |
r->es = seg; |
| 1431 |
r->fs = seg; |
| 1432 |
r->gs = seg; |
| 1433 |
r->eflags = VIF_MASK; |
| 1434 |
|
| 1435 |
/* move code to proper address. We use the same layout as a .com
|
| 1436 |
dos program. */
|
| 1437 |
memcpy(vm86_mem + (VM86_CODE_CS << 4) + VM86_CODE_IP,
|
| 1438 |
&vm86_code_start, &vm86_code_end - &vm86_code_start); |
| 1439 |
|
| 1440 |
/* mark int 0x21 as being emulated */
|
| 1441 |
set_bit((uint8_t *)&ctx.int_revectored, 0x21);
|
| 1442 |
|
| 1443 |
for(;;) {
|
| 1444 |
ret = vm86(VM86_ENTER, &ctx); |
| 1445 |
switch(VM86_TYPE(ret)) {
|
| 1446 |
case VM86_INTx:
|
| 1447 |
{
|
| 1448 |
int int_num, ah, v;
|
| 1449 |
|
| 1450 |
int_num = VM86_ARG(ret); |
| 1451 |
if (int_num != 0x21) |
| 1452 |
goto unknown_int;
|
| 1453 |
ah = (r->eax >> 8) & 0xff; |
| 1454 |
switch(ah) {
|
| 1455 |
case 0x00: /* exit */ |
| 1456 |
goto the_end;
|
| 1457 |
case 0x02: /* write char */ |
| 1458 |
{
|
| 1459 |
uint8_t c = r->edx; |
| 1460 |
putchar(c); |
| 1461 |
} |
| 1462 |
break;
|
| 1463 |
case 0x09: /* write string */ |
| 1464 |
{
|
| 1465 |
uint8_t c, *ptr; |
| 1466 |
ptr = seg_to_linear(r->ds, r->edx); |
| 1467 |
for(;;) {
|
| 1468 |
c = *ptr++; |
| 1469 |
if (c == '$') |
| 1470 |
break;
|
| 1471 |
putchar(c); |
| 1472 |
} |
| 1473 |
r->eax = (r->eax & ~0xff) | '$'; |
| 1474 |
} |
| 1475 |
break;
|
| 1476 |
case 0xff: /* extension: write eflags number in edx */ |
| 1477 |
v = (int)r->edx;
|
| 1478 |
#ifndef LINUX_VM86_IOPL_FIX
|
| 1479 |
v &= ~0x3000;
|
| 1480 |
#endif
|
| 1481 |
printf("%08x\n", v);
|
| 1482 |
break;
|
| 1483 |
default:
|
| 1484 |
unknown_int:
|
| 1485 |
printf("unsupported int 0x%02x\n", int_num);
|
| 1486 |
goto the_end;
|
| 1487 |
} |
| 1488 |
} |
| 1489 |
break;
|
| 1490 |
case VM86_SIGNAL:
|
| 1491 |
/* a signal came, we just ignore that */
|
| 1492 |
break;
|
| 1493 |
case VM86_STI:
|
| 1494 |
break;
|
| 1495 |
default:
|
| 1496 |
printf("ERROR: unhandled vm86 return code (0x%x)\n", ret);
|
| 1497 |
goto the_end;
|
| 1498 |
} |
| 1499 |
} |
| 1500 |
the_end:
|
| 1501 |
printf("VM86 end\n");
|
| 1502 |
munmap(vm86_mem, 0x110000);
|
| 1503 |
} |
| 1504 |
#endif
|
| 1505 |
|
| 1506 |
/* exception tests */
|
| 1507 |
#if defined(__i386__) && !defined(REG_EAX)
|
| 1508 |
#define REG_EAX EAX
|
| 1509 |
#define REG_EBX EBX
|
| 1510 |
#define REG_ECX ECX
|
| 1511 |
#define REG_EDX EDX
|
| 1512 |
#define REG_ESI ESI
|
| 1513 |
#define REG_EDI EDI
|
| 1514 |
#define REG_EBP EBP
|
| 1515 |
#define REG_ESP ESP
|
| 1516 |
#define REG_EIP EIP
|
| 1517 |
#define REG_EFL EFL
|
| 1518 |
#define REG_TRAPNO TRAPNO
|
| 1519 |
#define REG_ERR ERR
|
| 1520 |
#endif
|
| 1521 |
|
| 1522 |
#if defined(__x86_64__)
|
| 1523 |
#define REG_EIP REG_RIP
|
| 1524 |
#endif
|
| 1525 |
|
| 1526 |
jmp_buf jmp_env; |
| 1527 |
int v1;
|
| 1528 |
int tab[2]; |
| 1529 |
|
| 1530 |
void sig_handler(int sig, siginfo_t *info, void *puc) |
| 1531 |
{
|
| 1532 |
struct ucontext *uc = puc;
|
| 1533 |
|
| 1534 |
printf("si_signo=%d si_errno=%d si_code=%d",
|
| 1535 |
info->si_signo, info->si_errno, info->si_code); |
| 1536 |
printf(" si_addr=0x%08lx",
|
| 1537 |
(unsigned long)info->si_addr); |
| 1538 |
printf("\n");
|
| 1539 |
|
| 1540 |
printf("trapno=" FMTLX " err=" FMTLX, |
| 1541 |
(long)uc->uc_mcontext.gregs[REG_TRAPNO],
|
| 1542 |
(long)uc->uc_mcontext.gregs[REG_ERR]);
|
| 1543 |
printf(" EIP=" FMTLX, (long)uc->uc_mcontext.gregs[REG_EIP]); |
| 1544 |
printf("\n");
|
| 1545 |
longjmp(jmp_env, 1);
|
| 1546 |
} |
| 1547 |
|
| 1548 |
void test_exceptions(void) |
| 1549 |
{
|
| 1550 |
struct sigaction act;
|
| 1551 |
volatile int val; |
| 1552 |
|
| 1553 |
act.sa_sigaction = sig_handler; |
| 1554 |
sigemptyset(&act.sa_mask); |
| 1555 |
act.sa_flags = SA_SIGINFO | SA_NODEFER; |
| 1556 |
sigaction(SIGFPE, &act, NULL);
|
| 1557 |
sigaction(SIGILL, &act, NULL);
|
| 1558 |
sigaction(SIGSEGV, &act, NULL);
|
| 1559 |
sigaction(SIGBUS, &act, NULL);
|
| 1560 |
sigaction(SIGTRAP, &act, NULL);
|
| 1561 |
|
| 1562 |
/* test division by zero reporting */
|
| 1563 |
printf("DIVZ exception:\n");
|
| 1564 |
if (setjmp(jmp_env) == 0) { |
| 1565 |
/* now divide by zero */
|
| 1566 |
v1 = 0;
|
| 1567 |
v1 = 2 / v1;
|
| 1568 |
} |
| 1569 |
|
| 1570 |
#if !defined(__x86_64__)
|
| 1571 |
printf("BOUND exception:\n");
|
| 1572 |
if (setjmp(jmp_env) == 0) { |
| 1573 |
/* bound exception */
|
| 1574 |
tab[0] = 1; |
| 1575 |
tab[1] = 10; |
| 1576 |
asm volatile ("bound %0, %1" : : "r" (11), "m" (tab[0])); |
| 1577 |
} |
| 1578 |
#endif
|
| 1579 |
|
| 1580 |
#ifdef TEST_SEGS
|
| 1581 |
printf("segment exceptions:\n");
|
| 1582 |
if (setjmp(jmp_env) == 0) { |
| 1583 |
/* load an invalid segment */
|
| 1584 |
asm volatile ("movl %0, %%fs" : : "r" ((0x1234 << 3) | 1)); |
| 1585 |
} |
| 1586 |
if (setjmp(jmp_env) == 0) { |
| 1587 |
/* null data segment is valid */
|
| 1588 |
asm volatile ("movl %0, %%fs" : : "r" (3)); |
| 1589 |
/* null stack segment */
|
| 1590 |
asm volatile ("movl %0, %%ss" : : "r" (3)); |
| 1591 |
} |
| 1592 |
|
| 1593 |
{
|
| 1594 |
struct modify_ldt_ldt_s ldt;
|
| 1595 |
ldt.entry_number = 1;
|
| 1596 |
ldt.base_addr = (unsigned long)&seg_data1; |
| 1597 |
ldt.limit = (sizeof(seg_data1) + 0xfff) >> 12; |
| 1598 |
ldt.seg_32bit = 1;
|
| 1599 |
ldt.contents = MODIFY_LDT_CONTENTS_DATA; |
| 1600 |
ldt.read_exec_only = 0;
|
| 1601 |
ldt.limit_in_pages = 1;
|
| 1602 |
ldt.seg_not_present = 1;
|
| 1603 |
ldt.useable = 1;
|
| 1604 |
modify_ldt(1, &ldt, sizeof(ldt)); /* write ldt entry */ |
| 1605 |
|
| 1606 |
if (setjmp(jmp_env) == 0) { |
| 1607 |
/* segment not present */
|
| 1608 |
asm volatile ("movl %0, %%fs" : : "r" (MK_SEL(1))); |
| 1609 |
} |
| 1610 |
} |
| 1611 |
#endif
|
| 1612 |
|
| 1613 |
/* test SEGV reporting */
|
| 1614 |
printf("PF exception:\n");
|
| 1615 |
if (setjmp(jmp_env) == 0) { |
| 1616 |
val = 1;
|
| 1617 |
/* we add a nop to test a weird PC retrieval case */
|
| 1618 |
asm volatile ("nop"); |
| 1619 |
/* now store in an invalid address */
|
| 1620 |
*(char *)0x1234 = 1; |
| 1621 |
} |
| 1622 |
|
| 1623 |
/* test SEGV reporting */
|
| 1624 |
printf("PF exception:\n");
|
| 1625 |
if (setjmp(jmp_env) == 0) { |
| 1626 |
val = 1;
|
| 1627 |
/* read from an invalid address */
|
| 1628 |
v1 = *(char *)0x1234; |
| 1629 |
} |
| 1630 |
|
| 1631 |
/* test illegal instruction reporting */
|
| 1632 |
printf("UD2 exception:\n");
|
| 1633 |
if (setjmp(jmp_env) == 0) { |
| 1634 |
/* now execute an invalid instruction */
|
| 1635 |
asm volatile("ud2"); |
| 1636 |
} |
| 1637 |
printf("lock nop exception:\n");
|
| 1638 |
if (setjmp(jmp_env) == 0) { |
| 1639 |
/* now execute an invalid instruction */
|
| 1640 |
asm volatile("lock nop"); |
| 1641 |
} |
| 1642 |
|
| 1643 |
printf("INT exception:\n");
|
| 1644 |
if (setjmp(jmp_env) == 0) { |
| 1645 |
asm volatile ("int $0xfd"); |
| 1646 |
} |
| 1647 |
if (setjmp(jmp_env) == 0) { |
| 1648 |
asm volatile ("int $0x01"); |
| 1649 |
} |
| 1650 |
if (setjmp(jmp_env) == 0) { |
| 1651 |
asm volatile (".byte 0xcd, 0x03"); |
| 1652 |
} |
| 1653 |
if (setjmp(jmp_env) == 0) { |
| 1654 |
asm volatile ("int $0x04"); |
| 1655 |
} |
| 1656 |
if (setjmp(jmp_env) == 0) { |
| 1657 |
asm volatile ("int $0x05"); |
| 1658 |
} |
| 1659 |
|
| 1660 |
printf("INT3 exception:\n");
|
| 1661 |
if (setjmp(jmp_env) == 0) { |
| 1662 |
asm volatile ("int3"); |
| 1663 |
} |
| 1664 |
|
| 1665 |
printf("CLI exception:\n");
|
| 1666 |
if (setjmp(jmp_env) == 0) { |
| 1667 |
asm volatile ("cli"); |
| 1668 |
} |
| 1669 |
|
| 1670 |
printf("STI exception:\n");
|
| 1671 |
if (setjmp(jmp_env) == 0) { |
| 1672 |
asm volatile ("cli"); |
| 1673 |
} |
| 1674 |
|
| 1675 |
#if !defined(__x86_64__)
|
| 1676 |
printf("INTO exception:\n");
|
| 1677 |
if (setjmp(jmp_env) == 0) { |
| 1678 |
/* overflow exception */
|
| 1679 |
asm volatile ("addl $1, %0 ; into" : : "r" (0x7fffffff)); |
| 1680 |
} |
| 1681 |
#endif
|
| 1682 |
|
| 1683 |
printf("OUTB exception:\n");
|
| 1684 |
if (setjmp(jmp_env) == 0) { |
| 1685 |
asm volatile ("outb %%al, %%dx" : : "d" (0x4321), "a" (0)); |
| 1686 |
} |
| 1687 |
|
| 1688 |
printf("INB exception:\n");
|
| 1689 |
if (setjmp(jmp_env) == 0) { |
| 1690 |
asm volatile ("inb %%dx, %%al" : "=a" (val) : "d" (0x4321)); |
| 1691 |
} |
| 1692 |
|
| 1693 |
printf("REP OUTSB exception:\n");
|
| 1694 |
if (setjmp(jmp_env) == 0) { |
| 1695 |
asm volatile ("rep outsb" : : "d" (0x4321), "S" (tab), "c" (1)); |
| 1696 |
} |
| 1697 |
|
| 1698 |
printf("REP INSB exception:\n");
|
| 1699 |
if (setjmp(jmp_env) == 0) { |
| 1700 |
asm volatile ("rep insb" : : "d" (0x4321), "D" (tab), "c" (1)); |
| 1701 |
} |
| 1702 |
|
| 1703 |
printf("HLT exception:\n");
|
| 1704 |
if (setjmp(jmp_env) == 0) { |
| 1705 |
asm volatile ("hlt"); |
| 1706 |
} |
| 1707 |
|
| 1708 |
printf("single step exception:\n");
|
| 1709 |
val = 0;
|
| 1710 |
if (setjmp(jmp_env) == 0) { |
| 1711 |
asm volatile ("pushf\n" |
| 1712 |
"orl $0x00100, (%%esp)\n"
|
| 1713 |
"popf\n"
|
| 1714 |
"movl $0xabcd, %0\n"
|
| 1715 |
"movl $0x0, %0\n" : "=m" (val) : : "cc", "memory"); |
| 1716 |
} |
| 1717 |
printf("val=0x%x\n", val);
|
| 1718 |
} |
| 1719 |
|
| 1720 |
#if !defined(__x86_64__)
|
| 1721 |
/* specific precise single step test */
|
| 1722 |
void sig_trap_handler(int sig, siginfo_t *info, void *puc) |
| 1723 |
{
|
| 1724 |
struct ucontext *uc = puc;
|
| 1725 |
printf("EIP=" FMTLX "\n", (long)uc->uc_mcontext.gregs[REG_EIP]); |
| 1726 |
} |
| 1727 |
|
| 1728 |
const uint8_t sstep_buf1[4] = { 1, 2, 3, 4}; |
| 1729 |
uint8_t sstep_buf2[4];
|
| 1730 |
|
| 1731 |
void test_single_step(void) |
| 1732 |
{
|
| 1733 |
struct sigaction act;
|
| 1734 |
volatile int val; |
| 1735 |
int i;
|
| 1736 |
|
| 1737 |
val = 0;
|
| 1738 |
act.sa_sigaction = sig_trap_handler; |
| 1739 |
sigemptyset(&act.sa_mask); |
| 1740 |
act.sa_flags = SA_SIGINFO; |
| 1741 |
sigaction(SIGTRAP, &act, NULL);
|
| 1742 |
asm volatile ("pushf\n" |
| 1743 |
"orl $0x00100, (%%esp)\n"
|
| 1744 |
"popf\n"
|
| 1745 |
"movl $0xabcd, %0\n"
|
| 1746 |
|
| 1747 |
/* jmp test */
|
| 1748 |
"movl $3, %%ecx\n"
|
| 1749 |
"1:\n"
|
| 1750 |
"addl $1, %0\n"
|
| 1751 |
"decl %%ecx\n"
|
| 1752 |
"jnz 1b\n"
|
| 1753 |
|
| 1754 |
/* movsb: the single step should stop at each movsb iteration */
|
| 1755 |
"movl $sstep_buf1, %%esi\n"
|
| 1756 |
"movl $sstep_buf2, %%edi\n"
|
| 1757 |
"movl $0, %%ecx\n"
|
| 1758 |
"rep movsb\n"
|
| 1759 |
"movl $3, %%ecx\n"
|
| 1760 |
"rep movsb\n"
|
| 1761 |
"movl $1, %%ecx\n"
|
| 1762 |
"rep movsb\n"
|
| 1763 |
|
| 1764 |
/* cmpsb: the single step should stop at each cmpsb iteration */
|
| 1765 |
"movl $sstep_buf1, %%esi\n"
|
| 1766 |
"movl $sstep_buf2, %%edi\n"
|
| 1767 |
"movl $0, %%ecx\n"
|
| 1768 |
"rep cmpsb\n"
|
| 1769 |
"movl $4, %%ecx\n"
|
| 1770 |
"rep cmpsb\n"
|
| 1771 |
|
| 1772 |
/* getpid() syscall: single step should skip one
|
| 1773 |
instruction */
|
| 1774 |
"movl $20, %%eax\n"
|
| 1775 |
"int $0x80\n"
|
| 1776 |
"movl $0, %%eax\n"
|
| 1777 |
|
| 1778 |
/* when modifying SS, trace is not done on the next
|
| 1779 |
instruction */
|
| 1780 |
"movl %%ss, %%ecx\n"
|
| 1781 |
"movl %%ecx, %%ss\n"
|
| 1782 |
"addl $1, %0\n"
|
| 1783 |
"movl $1, %%eax\n"
|
| 1784 |
"movl %%ecx, %%ss\n"
|
| 1785 |
"jmp 1f\n"
|
| 1786 |
"addl $1, %0\n"
|
| 1787 |
"1:\n"
|
| 1788 |
"movl $1, %%eax\n"
|
| 1789 |
"pushl %%ecx\n"
|
| 1790 |
"popl %%ss\n"
|
| 1791 |
"addl $1, %0\n"
|
| 1792 |
"movl $1, %%eax\n"
|
| 1793 |
|
| 1794 |
"pushf\n"
|
| 1795 |
"andl $~0x00100, (%%esp)\n"
|
| 1796 |
"popf\n"
|
| 1797 |
: "=m" (val)
|
| 1798 |
: |
| 1799 |
: "cc", "memory", "eax", "ecx", "esi", "edi"); |
| 1800 |
printf("val=%d\n", val);
|
| 1801 |
for(i = 0; i < 4; i++) |
| 1802 |
printf("sstep_buf2[%d] = %d\n", i, sstep_buf2[i]);
|
| 1803 |
} |
| 1804 |
|
| 1805 |
/* self modifying code test */
|
| 1806 |
uint8_t code[] = {
|
| 1807 |
0xb8, 0x1, 0x00, 0x00, 0x00, /* movl $1, %eax */ |
| 1808 |
0xc3, /* ret */ |
| 1809 |
}; |
| 1810 |
|
| 1811 |
asm("smc_code2:\n" |
| 1812 |
"movl 4(%esp), %eax\n"
|
| 1813 |
"movl %eax, smc_patch_addr2 + 1\n"
|
| 1814 |
"nop\n"
|
| 1815 |
"nop\n"
|
| 1816 |
"nop\n"
|
| 1817 |
"nop\n"
|
| 1818 |
"nop\n"
|
| 1819 |
"nop\n"
|
| 1820 |
"nop\n"
|
| 1821 |
"nop\n"
|
| 1822 |
"smc_patch_addr2:\n"
|
| 1823 |
"movl $1, %eax\n"
|
| 1824 |
"ret\n");
|
| 1825 |
|
| 1826 |
typedef int FuncType(void); |
| 1827 |
extern int smc_code2(int); |
| 1828 |
void test_self_modifying_code(void) |
| 1829 |
{
|
| 1830 |
int i;
|
| 1831 |
|
| 1832 |
printf("self modifying code:\n");
|
| 1833 |
printf("func1 = 0x%x\n", ((FuncType *)code)());
|
| 1834 |
for(i = 2; i <= 4; i++) { |
| 1835 |
code[1] = i;
|
| 1836 |
printf("func%d = 0x%x\n", i, ((FuncType *)code)());
|
| 1837 |
} |
| 1838 |
|
| 1839 |
/* more difficult test : the modified code is just after the
|
| 1840 |
modifying instruction. It is forbidden in Intel specs, but it
|
| 1841 |
is used by old DOS programs */
|
| 1842 |
for(i = 2; i <= 4; i++) { |
| 1843 |
printf("smc_code2(%d) = %d\n", i, smc_code2(i));
|
| 1844 |
} |
| 1845 |
} |
| 1846 |
#endif
|
| 1847 |
|
| 1848 |
long enter_stack[4096]; |
| 1849 |
|
| 1850 |
#if defined(__x86_64__)
|
| 1851 |
#define RSP "%%rsp" |
| 1852 |
#define RBP "%%rbp" |
| 1853 |
#else
|
| 1854 |
#define RSP "%%esp" |
| 1855 |
#define RBP "%%ebp" |
| 1856 |
#endif
|
| 1857 |
|
| 1858 |
#define TEST_ENTER(size, stack_type, level)\
|
| 1859 |
{\
|
| 1860 |
long esp_save, esp_val, ebp_val, ebp_save, i;\
|
| 1861 |
stack_type *ptr, *stack_end, *stack_ptr;\ |
| 1862 |
memset(enter_stack, 0, sizeof(enter_stack));\ |
| 1863 |
stack_end = stack_ptr = (stack_type *)(enter_stack + 4096);\
|
| 1864 |
ebp_val = (long)stack_ptr;\
|
| 1865 |
for(i=1;i<=32;i++)\ |
| 1866 |
*--stack_ptr = i;\ |
| 1867 |
esp_val = (long)stack_ptr;\
|
| 1868 |
asm("mov " RSP ", %[esp_save]\n"\ |
| 1869 |
"mov " RBP ", %[ebp_save]\n"\ |
| 1870 |
"mov %[esp_val], " RSP "\n"\ |
| 1871 |
"mov %[ebp_val], " RBP "\n"\ |
| 1872 |
"enter" size " $8, $" #level "\n"\ |
| 1873 |
"mov " RSP ", %[esp_val]\n"\ |
| 1874 |
"mov " RBP ", %[ebp_val]\n"\ |
| 1875 |
"mov %[esp_save], " RSP "\n"\ |
| 1876 |
"mov %[ebp_save], " RBP "\n"\ |
| 1877 |
: [esp_save] "=r" (esp_save),\
|
| 1878 |
[ebp_save] "=r" (ebp_save),\
|
| 1879 |
[esp_val] "=r" (esp_val),\
|
| 1880 |
[ebp_val] "=r" (ebp_val)\
|
| 1881 |
: "[esp_val]" (esp_val),\
|
| 1882 |
"[ebp_val]" (ebp_val));\
|
| 1883 |
printf("level=%d:\n", level);\
|
| 1884 |
printf("esp_val=" FMTLX "\n", esp_val - (long)stack_end);\ |
| 1885 |
printf("ebp_val=" FMTLX "\n", ebp_val - (long)stack_end);\ |
| 1886 |
for(ptr = (stack_type *)esp_val; ptr < stack_end; ptr++)\
|
| 1887 |
printf(FMTLX "\n", (long)ptr[0]);\ |
| 1888 |
} |
| 1889 |
|
| 1890 |
static void test_enter(void) |
| 1891 |
{
|
| 1892 |
#if defined(__x86_64__)
|
| 1893 |
TEST_ENTER("q", uint64_t, 0); |
| 1894 |
TEST_ENTER("q", uint64_t, 1); |
| 1895 |
TEST_ENTER("q", uint64_t, 2); |
| 1896 |
TEST_ENTER("q", uint64_t, 31); |
| 1897 |
#else
|
| 1898 |
TEST_ENTER("l", uint32_t, 0); |
| 1899 |
TEST_ENTER("l", uint32_t, 1); |
| 1900 |
TEST_ENTER("l", uint32_t, 2); |
| 1901 |
TEST_ENTER("l", uint32_t, 31); |
| 1902 |
#endif
|
| 1903 |
|
| 1904 |
TEST_ENTER("w", uint16_t, 0); |
| 1905 |
TEST_ENTER("w", uint16_t, 1); |
| 1906 |
TEST_ENTER("w", uint16_t, 2); |
| 1907 |
TEST_ENTER("w", uint16_t, 31); |
| 1908 |
} |
| 1909 |
|
| 1910 |
#ifdef TEST_SSE
|
| 1911 |
|
| 1912 |
typedef int __m64 __attribute__ ((__mode__ (__V2SI__))); |
| 1913 |
typedef int __m128 __attribute__ ((__mode__(__V4SF__))); |
| 1914 |
|
| 1915 |
typedef union { |
| 1916 |
double d[2]; |
| 1917 |
float s[4]; |
| 1918 |
uint32_t l[4];
|
| 1919 |
uint64_t q[2];
|
| 1920 |
__m128 dq; |
| 1921 |
} XMMReg; |
| 1922 |
|
| 1923 |
static uint64_t __attribute__((aligned(16))) test_values[4][2] = { |
| 1924 |
{ 0x456723c698694873, 0xdc515cff944a58ec },
|
| 1925 |
{ 0x1f297ccd58bad7ab, 0x41f21efba9e3e146 },
|
| 1926 |
{ 0x007c62c2085427f8, 0x231be9e8cde7438d },
|
| 1927 |
{ 0x0f76255a085427f8, 0xc233e9e8c4c9439a },
|
| 1928 |
}; |
| 1929 |
|
| 1930 |
#define SSE_OP(op)\
|
| 1931 |
{\
|
| 1932 |
asm volatile (#op " %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\ |
| 1933 |
printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\ |
| 1934 |
#op,\
|
| 1935 |
a.q[1], a.q[0],\ |
| 1936 |
b.q[1], b.q[0],\ |
| 1937 |
r.q[1], r.q[0]);\ |
| 1938 |
} |
| 1939 |
|
| 1940 |
#define SSE_OP2(op)\
|
| 1941 |
{\
|
| 1942 |
int i;\
|
| 1943 |
for(i=0;i<2;i++) {\ |
| 1944 |
a.q[0] = test_values[2*i][0];\ |
| 1945 |
a.q[1] = test_values[2*i][1];\ |
| 1946 |
b.q[0] = test_values[2*i+1][0];\ |
| 1947 |
b.q[1] = test_values[2*i+1][1];\ |
| 1948 |
SSE_OP(op);\ |
| 1949 |
}\ |
| 1950 |
} |
| 1951 |
|
| 1952 |
#define MMX_OP2(op)\
|
| 1953 |
{\
|
| 1954 |
int i;\
|
| 1955 |
for(i=0;i<2;i++) {\ |
| 1956 |
a.q[0] = test_values[2*i][0];\ |
| 1957 |
b.q[0] = test_values[2*i+1][0];\ |
| 1958 |
asm volatile (#op " %2, %0" : "=y" (r.q[0]) : "0" (a.q[0]), "y" (b.q[0]));\ |
| 1959 |
printf("%-9s: a=" FMT64X " b=" FMT64X " r=" FMT64X "\n",\ |
| 1960 |
#op,\
|
| 1961 |
a.q[0],\
|
| 1962 |
b.q[0],\
|
| 1963 |
r.q[0]);\
|
| 1964 |
}\ |
| 1965 |
SSE_OP2(op);\ |
| 1966 |
} |
| 1967 |
|
| 1968 |
#define SHUF_OP(op, ib)\
|
| 1969 |
{\
|
| 1970 |
a.q[0] = test_values[0][0];\ |
| 1971 |
a.q[1] = test_values[0][1];\ |
| 1972 |
b.q[0] = test_values[1][0];\ |
| 1973 |
b.q[1] = test_values[1][1];\ |
| 1974 |
asm volatile (#op " $" #ib ", %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\ |
| 1975 |
printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\ |
| 1976 |
#op,\
|
| 1977 |
a.q[1], a.q[0],\ |
| 1978 |
b.q[1], b.q[0],\ |
| 1979 |
ib,\ |
| 1980 |
r.q[1], r.q[0]);\ |
| 1981 |
} |
| 1982 |
|
| 1983 |
#define PSHUF_OP(op, ib)\
|
| 1984 |
{\
|
| 1985 |
int i;\
|
| 1986 |
for(i=0;i<2;i++) {\ |
| 1987 |
a.q[0] = test_values[2*i][0];\ |
| 1988 |
a.q[1] = test_values[2*i][1];\ |
| 1989 |
asm volatile (#op " $" #ib ", %1, %0" : "=x" (r.dq) : "x" (a.dq));\ |
| 1990 |
printf("%-9s: a=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\ |
| 1991 |
#op,\
|
| 1992 |
a.q[1], a.q[0],\ |
| 1993 |
ib,\ |
| 1994 |
r.q[1], r.q[0]);\ |
| 1995 |
}\ |
| 1996 |
} |
| 1997 |
|
| 1998 |
#define SHIFT_IM(op, ib)\
|
| 1999 |
{\
|
| 2000 |
int i;\
|
| 2001 |
for(i=0;i<2;i++) {\ |
| 2002 |
a.q[0] = test_values[2*i][0];\ |
| 2003 |
a.q[1] = test_values[2*i][1];\ |
| 2004 |
asm volatile (#op " $" #ib ", %0" : "=x" (r.dq) : "0" (a.dq));\ |
| 2005 |
printf("%-9s: a=" FMT64X "" FMT64X " ib=%02x r=" FMT64X "" FMT64X "\n",\ |
| 2006 |
#op,\
|
| 2007 |
a.q[1], a.q[0],\ |
| 2008 |
ib,\ |
| 2009 |
r.q[1], r.q[0]);\ |
| 2010 |
}\ |
| 2011 |
} |
| 2012 |
|
| 2013 |
#define SHIFT_OP(op, ib)\
|
| 2014 |
{\
|
| 2015 |
int i;\
|
| 2016 |
SHIFT_IM(op, ib);\ |
| 2017 |
for(i=0;i<2;i++) {\ |
| 2018 |
a.q[0] = test_values[2*i][0];\ |
| 2019 |
a.q[1] = test_values[2*i][1];\ |
| 2020 |
b.q[0] = ib;\
|
| 2021 |
b.q[1] = 0;\ |
| 2022 |
asm volatile (#op " %2, %0" : "=x" (r.dq) : "0" (a.dq), "x" (b.dq));\ |
| 2023 |
printf("%-9s: a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\ |
| 2024 |
#op,\
|
| 2025 |
a.q[1], a.q[0],\ |
| 2026 |
b.q[1], b.q[0],\ |
| 2027 |
r.q[1], r.q[0]);\ |
| 2028 |
}\ |
| 2029 |
} |
| 2030 |
|
| 2031 |
#define MOVMSK(op)\
|
| 2032 |
{\
|
| 2033 |
int i, reg;\
|
| 2034 |
for(i=0;i<2;i++) {\ |
| 2035 |
a.q[0] = test_values[2*i][0];\ |
| 2036 |
a.q[1] = test_values[2*i][1];\ |
| 2037 |
asm volatile (#op " %1, %0" : "=r" (reg) : "x" (a.dq));\ |
| 2038 |
printf("%-9s: a=" FMT64X "" FMT64X " r=%08x\n",\ |
| 2039 |
#op,\
|
| 2040 |
a.q[1], a.q[0],\ |
| 2041 |
reg);\ |
| 2042 |
}\ |
| 2043 |
} |
| 2044 |
|
| 2045 |
#define SSE_OPS(a) \
|
| 2046 |
SSE_OP(a ## ps);\ |
| 2047 |
SSE_OP(a ## ss); |
| 2048 |
|
| 2049 |
#define SSE_OPD(a) \
|
| 2050 |
SSE_OP(a ## pd);\ |
| 2051 |
SSE_OP(a ## sd); |
| 2052 |
|
| 2053 |
#define SSE_COMI(op, field)\
|
| 2054 |
{\
|
| 2055 |
unsigned int eflags;\ |
| 2056 |
XMMReg a, b;\ |
| 2057 |
a.field[0] = a1;\
|
| 2058 |
b.field[0] = b1;\
|
| 2059 |
asm volatile (#op " %2, %1\n"\ |
| 2060 |
"pushf\n"\
|
| 2061 |
"pop %0\n"\
|
| 2062 |
: "=m" (eflags)\
|
| 2063 |
: "x" (a.dq), "x" (b.dq));\ |
| 2064 |
printf("%-9s: a=%f b=%f cc=%04x\n",\
|
| 2065 |
#op, a1, b1,\
|
| 2066 |
eflags & (CC_C | CC_P | CC_Z | CC_S | CC_O | CC_A));\ |
| 2067 |
} |
| 2068 |
|
| 2069 |
void test_sse_comi(double a1, double b1) |
| 2070 |
{
|
| 2071 |
SSE_COMI(ucomiss, s); |
| 2072 |
SSE_COMI(ucomisd, d); |
| 2073 |
SSE_COMI(comiss, s); |
| 2074 |
SSE_COMI(comisd, d); |
| 2075 |
} |
| 2076 |
|
| 2077 |
#define CVT_OP_XMM(op)\
|
| 2078 |
{\
|
| 2079 |
asm volatile (#op " %1, %0" : "=x" (r.dq) : "x" (a.dq));\ |
| 2080 |
printf("%-9s: a=" FMT64X "" FMT64X " r=" FMT64X "" FMT64X "\n",\ |
| 2081 |
#op,\
|
| 2082 |
a.q[1], a.q[0],\ |
| 2083 |
r.q[1], r.q[0]);\ |
| 2084 |
} |
| 2085 |
|
| 2086 |
#define CVT_OP_XMM2MMX(op)\
|
| 2087 |
{\
|
| 2088 |
asm volatile (#op " %1, %0" : "=y" (r.q[0]) : "x" (a.dq));\ |
| 2089 |
printf("%-9s: a=" FMT64X "" FMT64X " r=" FMT64X "\n",\ |
| 2090 |
#op,\
|
| 2091 |
a.q[1], a.q[0],\ |
| 2092 |
r.q[0]);\
|
| 2093 |
} |
| 2094 |
|
| 2095 |
#define CVT_OP_MMX2XMM(op)\
|
| 2096 |
{\
|
| 2097 |
asm volatile (#op " %1, %0" : "=x" (r.dq) : "y" (a.q[0]));\ |
| 2098 |
printf("%-9s: a=" FMT64X " r=" FMT64X "" FMT64X "\n",\ |
| 2099 |
#op,\
|
| 2100 |
a.q[0],\
|
| 2101 |
r.q[1], r.q[0]);\ |
| 2102 |
} |
| 2103 |
|
| 2104 |
#define CVT_OP_REG2XMM(op)\
|
| 2105 |
{\
|
| 2106 |
asm volatile (#op " %1, %0" : "=x" (r.dq) : "r" (a.l[0]));\ |
| 2107 |
printf("%-9s: a=%08x r=" FMT64X "" FMT64X "\n",\ |
| 2108 |
#op,\
|
| 2109 |
a.l[0],\
|
| 2110 |
r.q[1], r.q[0]);\ |
| 2111 |
} |
| 2112 |
|
| 2113 |
#define CVT_OP_XMM2REG(op)\
|
| 2114 |
{\
|
| 2115 |
asm volatile (#op " %1, %0" : "=r" (r.l[0]) : "x" (a.dq));\ |
| 2116 |
printf("%-9s: a=" FMT64X "" FMT64X " r=%08x\n",\ |
| 2117 |
#op,\
|
| 2118 |
a.q[1], a.q[0],\ |
| 2119 |
r.l[0]);\
|
| 2120 |
} |
| 2121 |
|
| 2122 |
struct fpxstate {
|
| 2123 |
uint16_t fpuc; |
| 2124 |
uint16_t fpus; |
| 2125 |
uint16_t fptag; |
| 2126 |
uint16_t fop; |
| 2127 |
uint32_t fpuip; |
| 2128 |
uint16_t cs_sel; |
| 2129 |
uint16_t dummy0; |
| 2130 |
uint32_t fpudp; |
| 2131 |
uint16_t ds_sel; |
| 2132 |
uint16_t dummy1; |
| 2133 |
uint32_t mxcsr; |
| 2134 |
uint32_t mxcsr_mask; |
| 2135 |
uint8_t fpregs1[8 * 16]; |
| 2136 |
uint8_t xmm_regs[8 * 16]; |
| 2137 |
uint8_t dummy2[224];
|
| 2138 |
}; |
| 2139 |
|
| 2140 |
static struct fpxstate fpx_state __attribute__((aligned(16))); |
| 2141 |
static struct fpxstate fpx_state2 __attribute__((aligned(16))); |
| 2142 |
|
| 2143 |
void test_fxsave(void) |
| 2144 |
{
|
| 2145 |
struct fpxstate *fp = &fpx_state;
|
| 2146 |
struct fpxstate *fp2 = &fpx_state2;
|
| 2147 |
int i, nb_xmm;
|
| 2148 |
XMMReg a, b; |
| 2149 |
a.q[0] = test_values[0][0]; |
| 2150 |
a.q[1] = test_values[0][1]; |
| 2151 |
b.q[0] = test_values[1][0]; |
| 2152 |
b.q[1] = test_values[1][1]; |
| 2153 |
|
| 2154 |
asm("movdqa %2, %%xmm0\n" |
| 2155 |
"movdqa %3, %%xmm7\n"
|
| 2156 |
#if defined(__x86_64__)
|
| 2157 |
"movdqa %2, %%xmm15\n"
|
| 2158 |
#endif
|
| 2159 |
" fld1\n"
|
| 2160 |
" fldpi\n"
|
| 2161 |
" fldln2\n"
|
| 2162 |
" fxsave %0\n"
|
| 2163 |
" fxrstor %0\n"
|
| 2164 |
" fxsave %1\n"
|
| 2165 |
" fninit\n"
|
| 2166 |
: "=m" (*(uint32_t *)fp2), "=m" (*(uint32_t *)fp) |
| 2167 |
: "m" (a), "m" (b)); |
| 2168 |
printf("fpuc=%04x\n", fp->fpuc);
|
| 2169 |
printf("fpus=%04x\n", fp->fpus);
|
| 2170 |
printf("fptag=%04x\n", fp->fptag);
|
| 2171 |
for(i = 0; i < 3; i++) { |
| 2172 |
printf("ST%d: " FMT64X " %04x\n", |
| 2173 |
i, |
| 2174 |
*(uint64_t *)&fp->fpregs1[i * 16],
|
| 2175 |
*(uint16_t *)&fp->fpregs1[i * 16 + 8]); |
| 2176 |
} |
| 2177 |
printf("mxcsr=%08x\n", fp->mxcsr & 0x1f80); |
| 2178 |
#if defined(__x86_64__)
|
| 2179 |
nb_xmm = 16;
|
| 2180 |
#else
|
| 2181 |
nb_xmm = 8;
|
| 2182 |
#endif
|
| 2183 |
for(i = 0; i < nb_xmm; i++) { |
| 2184 |
printf("xmm%d: " FMT64X "" FMT64X "\n", |
| 2185 |
i, |
| 2186 |
*(uint64_t *)&fp->xmm_regs[i * 16],
|
| 2187 |
*(uint64_t *)&fp->xmm_regs[i * 16 + 8]); |
| 2188 |
} |
| 2189 |
} |
| 2190 |
|
| 2191 |
void test_sse(void) |
| 2192 |
{
|
| 2193 |
XMMReg r, a, b; |
| 2194 |
|
| 2195 |
MMX_OP2(punpcklbw); |
| 2196 |
MMX_OP2(punpcklwd); |
| 2197 |
MMX_OP2(punpckldq); |
| 2198 |
MMX_OP2(packsswb); |
| 2199 |
MMX_OP2(pcmpgtb); |
| 2200 |
MMX_OP2(pcmpgtw); |
| 2201 |
MMX_OP2(pcmpgtd); |
| 2202 |
MMX_OP2(packuswb); |
| 2203 |
MMX_OP2(punpckhbw); |
| 2204 |
MMX_OP2(punpckhwd); |
| 2205 |
MMX_OP2(punpckhdq); |
| 2206 |
MMX_OP2(packssdw); |
| 2207 |
MMX_OP2(pcmpeqb); |
| 2208 |
MMX_OP2(pcmpeqw); |
| 2209 |
MMX_OP2(pcmpeqd); |
| 2210 |
|
| 2211 |
MMX_OP2(paddq); |
| 2212 |
MMX_OP2(pmullw); |
| 2213 |
MMX_OP2(psubusb); |
| 2214 |
MMX_OP2(psubusw); |
| 2215 |
MMX_OP2(pminub); |
| 2216 |
MMX_OP2(pand); |
| 2217 |
MMX_OP2(paddusb); |
| 2218 |
MMX_OP2(paddusw); |
| 2219 |
MMX_OP2(pmaxub); |
| 2220 |
MMX_OP2(pandn); |
| 2221 |
|
| 2222 |
MMX_OP2(pmulhuw); |
| 2223 |
MMX_OP2(pmulhw); |
| 2224 |
|
| 2225 |
MMX_OP2(psubsb); |
| 2226 |
MMX_OP2(psubsw); |
| 2227 |
MMX_OP2(pminsw); |
| 2228 |
MMX_OP2(por); |
| 2229 |
MMX_OP2(paddsb); |
| 2230 |
MMX_OP2(paddsw); |
| 2231 |
MMX_OP2(pmaxsw); |
| 2232 |
MMX_OP2(pxor); |
| 2233 |
MMX_OP2(pmuludq); |
| 2234 |
MMX_OP2(pmaddwd); |
| 2235 |
MMX_OP2(psadbw); |
| 2236 |
MMX_OP2(psubb); |
| 2237 |
MMX_OP2(psubw); |
| 2238 |
MMX_OP2(psubd); |
| 2239 |
MMX_OP2(psubq); |
| 2240 |
MMX_OP2(paddb); |
| 2241 |
MMX_OP2(paddw); |
| 2242 |
MMX_OP2(paddd); |
| 2243 |
|
| 2244 |
MMX_OP2(pavgb); |
| 2245 |
MMX_OP2(pavgw); |
| 2246 |
|
| 2247 |
asm volatile ("pinsrw $1, %1, %0" : "=y" (r.q[0]) : "r" (0x12345678)); |
| 2248 |
printf("%-9s: r=" FMT64X "\n", "pinsrw", r.q[0]); |
| 2249 |
|
| 2250 |
asm volatile ("pinsrw $5, %1, %0" : "=x" (r.dq) : "r" (0x12345678)); |
| 2251 |
printf("%-9s: r=" FMT64X "" FMT64X "\n", "pinsrw", r.q[1], r.q[0]); |
| 2252 |
|
| 2253 |
a.q[0] = test_values[0][0]; |
| 2254 |
a.q[1] = test_values[0][1]; |
| 2255 |
asm volatile ("pextrw $1, %1, %0" : "=r" (r.l[0]) : "y" (a.q[0])); |
| 2256 |
printf("%-9s: r=%08x\n", "pextrw", r.l[0]); |
| 2257 |
|
| 2258 |
asm volatile ("pextrw $5, %1, %0" : "=r" (r.l[0]) : "x" (a.dq)); |
| 2259 |
printf("%-9s: r=%08x\n", "pextrw", r.l[0]); |
| 2260 |
|
| 2261 |
asm volatile ("pmovmskb %1, %0" : "=r" (r.l[0]) : "y" (a.q[0])); |
| 2262 |
printf("%-9s: r=%08x\n", "pmovmskb", r.l[0]); |
| 2263 |
|
| 2264 |
asm volatile ("pmovmskb %1, %0" : "=r" (r.l[0]) : "x" (a.dq)); |
| 2265 |
printf("%-9s: r=%08x\n", "pmovmskb", r.l[0]); |
| 2266 |
|
| 2267 |
{
|
| 2268 |
r.q[0] = -1; |
| 2269 |
r.q[1] = -1; |
| 2270 |
|
| 2271 |
a.q[0] = test_values[0][0]; |
| 2272 |
a.q[1] = test_values[0][1]; |
| 2273 |
b.q[0] = test_values[1][0]; |
| 2274 |
b.q[1] = test_values[1][1]; |
| 2275 |
asm volatile("maskmovq %1, %0" : |
| 2276 |
: "y" (a.q[0]), "y" (b.q[0]), "D" (&r) |
| 2277 |
: "memory");
|
| 2278 |
printf("%-9s: r=" FMT64X " a=" FMT64X " b=" FMT64X "\n", |
| 2279 |
"maskmov",
|
| 2280 |
r.q[0],
|
| 2281 |
a.q[0],
|
| 2282 |
b.q[0]);
|
| 2283 |
asm volatile("maskmovdqu %1, %0" : |
| 2284 |
: "x" (a.dq), "x" (b.dq), "D" (&r) |
| 2285 |
: "memory");
|
| 2286 |
printf("%-9s: r=" FMT64X "" FMT64X " a=" FMT64X "" FMT64X " b=" FMT64X "" FMT64X "\n", |
| 2287 |
"maskmov",
|
| 2288 |
r.q[1], r.q[0], |
| 2289 |
a.q[1], a.q[0], |
| 2290 |
b.q[1], b.q[0]); |
| 2291 |
} |
| 2292 |
|
| 2293 |
asm volatile ("emms"); |
| 2294 |
|
| 2295 |
SSE_OP2(punpcklqdq); |
| 2296 |
SSE_OP2(punpckhqdq); |
| 2297 |
SSE_OP2(andps); |
| 2298 |
SSE_OP2(andpd); |
| 2299 |
SSE_OP2(andnps); |
| 2300 |
SSE_OP2(andnpd); |
| 2301 |
SSE_OP2(orps); |
| 2302 |
SSE_OP2(orpd); |
| 2303 |
SSE_OP2(xorps); |
| 2304 |
SSE_OP2(xorpd); |
| 2305 |
|
| 2306 |
SSE_OP2(unpcklps); |
| 2307 |
SSE_OP2(unpcklpd); |
| 2308 |
SSE_OP2(unpckhps); |
| 2309 |
SSE_OP2(unpckhpd); |
| 2310 |
|
| 2311 |
SHUF_OP(shufps, 0x78);
|
| 2312 |
SHUF_OP(shufpd, 0x02);
|
| 2313 |
|
| 2314 |
PSHUF_OP(pshufd, 0x78);
|
| 2315 |
PSHUF_OP(pshuflw, 0x78);
|
| 2316 |
PSHUF_OP(pshufhw, 0x78);
|
| 2317 |
|
| 2318 |
SHIFT_OP(psrlw, 7);
|
| 2319 |
SHIFT_OP(psrlw, 16);
|
| 2320 |
SHIFT_OP(psraw, 7);
|
| 2321 |
SHIFT_OP(psraw, 16);
|
| 2322 |
SHIFT_OP(psllw, 7);
|
| 2323 |
SHIFT_OP(psllw, 16);
|
| 2324 |
|
| 2325 |
SHIFT_OP(psrld, 7);
|
| 2326 |
SHIFT_OP(psrld, 32);
|
| 2327 |
SHIFT_OP(psrad, 7);
|
| 2328 |
SHIFT_OP(psrad, 32);
|
| 2329 |
SHIFT_OP(pslld, 7);
|
| 2330 |
SHIFT_OP(pslld, 32);
|
| 2331 |
|
| 2332 |
SHIFT_OP(psrlq, 7);
|
| 2333 |
SHIFT_OP(psrlq, 32);
|
| 2334 |
SHIFT_OP(psllq, 7);
|
| 2335 |
SHIFT_OP(psllq, 32);
|
| 2336 |
|
| 2337 |
SHIFT_IM(psrldq, 16);
|
| 2338 |
SHIFT_IM(psrldq, 7);
|
| 2339 |
SHIFT_IM(pslldq, 16);
|
| 2340 |
SHIFT_IM(pslldq, 7);
|
| 2341 |
|
| 2342 |
MOVMSK(movmskps); |
| 2343 |
MOVMSK(movmskpd); |
| 2344 |
|
| 2345 |
/* FPU specific ops */
|
| 2346 |
|
| 2347 |
{
|
| 2348 |
uint32_t mxcsr; |
| 2349 |
asm volatile("stmxcsr %0" : "=m" (mxcsr)); |
| 2350 |
printf("mxcsr=%08x\n", mxcsr & 0x1f80); |
| 2351 |
asm volatile("ldmxcsr %0" : : "m" (mxcsr)); |
| 2352 |
} |
| 2353 |
|
| 2354 |
test_sse_comi(2, -1); |
| 2355 |
test_sse_comi(2, 2); |
| 2356 |
test_sse_comi(2, 3); |
| 2357 |
|
| 2358 |
a.s[0] = 2.7; |
| 2359 |
a.s[1] = 3.4; |
| 2360 |
a.s[2] = 4; |
| 2361 |
a.s[3] = -6.3; |
| 2362 |
b.s[0] = 45.7; |
| 2363 |
b.s[1] = 353.4; |
| 2364 |
b.s[2] = 4; |
| 2365 |
b.s[3] = 56.3; |
| 2366 |
SSE_OPS(add); |
| 2367 |
SSE_OPS(mul); |
| 2368 |
SSE_OPS(sub); |
| 2369 |
SSE_OPS(min); |
| 2370 |
SSE_OPS(div); |
| 2371 |
SSE_OPS(max); |
| 2372 |
SSE_OPS(sqrt); |
| 2373 |
SSE_OPS(cmpeq); |
| 2374 |
SSE_OPS(cmplt); |
| 2375 |
SSE_OPS(cmple); |
| 2376 |
SSE_OPS(cmpunord); |
| 2377 |
SSE_OPS(cmpneq); |
| 2378 |
SSE_OPS(cmpnlt); |
| 2379 |
SSE_OPS(cmpnle); |
| 2380 |
SSE_OPS(cmpord); |
| 2381 |
|
| 2382 |
a.d[0] = 2.7; |
| 2383 |
a.d[1] = -3.4; |
| 2384 |
b.d[0] = 45.7; |
| 2385 |
b.d[1] = -53.4; |
| 2386 |
SSE_OPD(add); |
| 2387 |
SSE_OPD(mul); |
| 2388 |
SSE_OPD(sub); |
| 2389 |
SSE_OPD(min); |
| 2390 |
SSE_OPD(div); |
| 2391 |
SSE_OPD(max); |
| 2392 |
SSE_OPD(sqrt); |
| 2393 |
SSE_OPD(cmpeq); |
| 2394 |
SSE_OPD(cmplt); |
| 2395 |
SSE_OPD(cmple); |
| 2396 |
SSE_OPD(cmpunord); |
| 2397 |
SSE_OPD(cmpneq); |
| 2398 |
SSE_OPD(cmpnlt); |
| 2399 |
SSE_OPD(cmpnle); |
| 2400 |
SSE_OPD(cmpord); |
| 2401 |
|
| 2402 |
/* float to float/int */
|
| 2403 |
a.s[0] = 2.7; |
| 2404 |
a.s[1] = 3.4; |
| 2405 |
a.s[2] = 4; |
| 2406 |
a.s[3] = -6.3; |
| 2407 |
CVT_OP_XMM(cvtps2pd); |
| 2408 |
CVT_OP_XMM(cvtss2sd); |
| 2409 |
CVT_OP_XMM2MMX(cvtps2pi); |
| 2410 |
CVT_OP_XMM2MMX(cvttps2pi); |
| 2411 |
CVT_OP_XMM2REG(cvtss2si); |
| 2412 |
CVT_OP_XMM2REG(cvttss2si); |
| 2413 |
CVT_OP_XMM(cvtps2dq); |
| 2414 |
CVT_OP_XMM(cvttps2dq); |
| 2415 |
|
| 2416 |
a.d[0] = 2.6; |
| 2417 |
a.d[1] = -3.4; |
| 2418 |
CVT_OP_XMM(cvtpd2ps); |
| 2419 |
CVT_OP_XMM(cvtsd2ss); |
| 2420 |
CVT_OP_XMM2MMX(cvtpd2pi); |
| 2421 |
CVT_OP_XMM2MMX(cvttpd2pi); |
| 2422 |
CVT_OP_XMM2REG(cvtsd2si); |
| 2423 |
CVT_OP_XMM2REG(cvttsd2si); |
| 2424 |
CVT_OP_XMM(cvtpd2dq); |
| 2425 |
CVT_OP_XMM(cvttpd2dq); |
| 2426 |
|
| 2427 |
/* int to float */
|
| 2428 |
a.l[0] = -6; |
| 2429 |
a.l[1] = 2; |
| 2430 |
a.l[2] = 100; |
| 2431 |
a.l[3] = -60000; |
| 2432 |
CVT_OP_MMX2XMM(cvtpi2ps); |
| 2433 |
CVT_OP_MMX2XMM(cvtpi2pd); |
| 2434 |
CVT_OP_REG2XMM(cvtsi2ss); |
| 2435 |
CVT_OP_REG2XMM(cvtsi2sd); |
| 2436 |
CVT_OP_XMM(cvtdq2ps); |
| 2437 |
CVT_OP_XMM(cvtdq2pd); |
| 2438 |
|
| 2439 |
/* XXX: test PNI insns */
|
| 2440 |
#if 0
|
| 2441 |
SSE_OP2(movshdup);
|
| 2442 |
#endif
|
| 2443 |
asm volatile ("emms"); |
| 2444 |
} |
| 2445 |
|
| 2446 |
#endif
|
| 2447 |
|
| 2448 |
extern void *__start_initcall; |
| 2449 |
extern void *__stop_initcall; |
| 2450 |
|
| 2451 |
|
| 2452 |
int main(int argc, char **argv) |
| 2453 |
{
|
| 2454 |
void **ptr;
|
| 2455 |
void (*func)(void); |
| 2456 |
|
| 2457 |
ptr = &__start_initcall; |
| 2458 |
while (ptr != &__stop_initcall) {
|
| 2459 |
func = *ptr++; |
| 2460 |
func(); |
| 2461 |
} |
| 2462 |
test_bsx(); |
| 2463 |
test_mul(); |
| 2464 |
test_jcc(); |
| 2465 |
test_floats(); |
| 2466 |
#if !defined(__x86_64__)
|
| 2467 |
test_bcd(); |
| 2468 |
#endif
|
| 2469 |
test_xchg(); |
| 2470 |
test_string(); |
| 2471 |
test_misc(); |
| 2472 |
test_lea(); |
| 2473 |
#ifdef TEST_SEGS
|
| 2474 |
test_segs(); |
| 2475 |
test_code16(); |
| 2476 |
#endif
|
| 2477 |
#ifdef TEST_VM86
|
| 2478 |
test_vm86(); |
| 2479 |
#endif
|
| 2480 |
test_exceptions(); |
| 2481 |
#if !defined(__x86_64__)
|
| 2482 |
test_self_modifying_code(); |
| 2483 |
test_single_step(); |
| 2484 |
#endif
|
| 2485 |
test_enter(); |
| 2486 |
#ifdef TEST_SSE
|
| 2487 |
test_sse(); |
| 2488 |
test_fxsave(); |
| 2489 |
#endif
|
| 2490 |
return 0; |
| 2491 |
} |