| 1 | /* $Id: tstInlineAsm.cpp 25491 2009-12-18 15:20:48Z vboxsync $ */
|
|---|
| 2 | /** @file
|
|---|
| 3 | * IPRT Testcase - inline assembly.
|
|---|
| 4 | */
|
|---|
| 5 |
|
|---|
| 6 | /*
|
|---|
| 7 | * Copyright (C) 2006-2007 Sun Microsystems, Inc.
|
|---|
| 8 | *
|
|---|
| 9 | * This file is part of VirtualBox Open Source Edition (OSE), as
|
|---|
| 10 | * available from http://www.virtualbox.org. This file is free software;
|
|---|
| 11 | * you can redistribute it and/or modify it under the terms of the GNU
|
|---|
| 12 | * General Public License (GPL) as published by the Free Software
|
|---|
| 13 | * Foundation, in version 2 as it comes in the "COPYING" file of the
|
|---|
| 14 | * VirtualBox OSE distribution. VirtualBox OSE is distributed in the
|
|---|
| 15 | * hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
|
|---|
| 16 | *
|
|---|
| 17 | * The contents of this file may alternatively be used under the terms
|
|---|
| 18 | * of the Common Development and Distribution License Version 1.0
|
|---|
| 19 | * (CDDL) only, as it comes in the "COPYING.CDDL" file of the
|
|---|
| 20 | * VirtualBox OSE distribution, in which case the provisions of the
|
|---|
| 21 | * CDDL are applicable instead of those of the GPL.
|
|---|
| 22 | *
|
|---|
| 23 | * You may elect to license modified versions of this file under the
|
|---|
| 24 | * terms and conditions of either the GPL or the CDDL or both.
|
|---|
| 25 | *
|
|---|
| 26 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa
|
|---|
| 27 | * Clara, CA 95054 USA or visit http://www.sun.com if you need
|
|---|
| 28 | * additional information or have any questions.
|
|---|
| 29 | */
|
|---|
| 30 |
|
|---|
| 31 | /*******************************************************************************
|
|---|
| 32 | * Header Files *
|
|---|
| 33 | *******************************************************************************/
|
|---|
| 34 | #include <iprt/asm.h>
|
|---|
| 35 | #include <iprt/stream.h>
|
|---|
| 36 | #include <iprt/string.h>
|
|---|
| 37 | #include <iprt/initterm.h>
|
|---|
| 38 | #include <iprt/param.h>
|
|---|
| 39 | #include <iprt/thread.h>
|
|---|
| 40 | #include <iprt/test.h>
|
|---|
| 41 |
|
|---|
| 42 |
|
|---|
| 43 |
|
|---|
| 44 | /*******************************************************************************
|
|---|
| 45 | * Defined Constants And Macros *
|
|---|
| 46 | *******************************************************************************/
|
|---|
| 47 | #define CHECKVAL(val, expect, fmt) \
|
|---|
| 48 | do \
|
|---|
| 49 | { \
|
|---|
| 50 | if ((val) != (expect)) \
|
|---|
| 51 | { \
|
|---|
| 52 | RTTestIErrorInc(); \
|
|---|
| 53 | RTPrintf("%s, %d: " #val ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (expect), (val)); \
|
|---|
| 54 | } \
|
|---|
| 55 | } while (0)
|
|---|
| 56 |
|
|---|
| 57 | #define CHECKOP(op, expect, fmt, type) \
|
|---|
| 58 | do \
|
|---|
| 59 | { \
|
|---|
| 60 | type val = op; \
|
|---|
| 61 | if (val != (type)(expect)) \
|
|---|
| 62 | { \
|
|---|
| 63 | RTTestIErrorInc(); \
|
|---|
| 64 | RTPrintf("%s, %d: " #op ": expected " fmt " got " fmt "\n", __FUNCTION__, __LINE__, (type)(expect), val); \
|
|---|
| 65 | } \
|
|---|
| 66 | } while (0)
|
|---|
| 67 |
|
|---|
| 68 |
|
|---|
| 69 | #if !defined(PIC) || !defined(RT_ARCH_X86)
|
|---|
| 70 | const char *getCacheAss(unsigned u)
|
|---|
| 71 | {
|
|---|
| 72 | if (u == 0)
|
|---|
| 73 | return "res0 ";
|
|---|
| 74 | if (u == 1)
|
|---|
| 75 | return "direct";
|
|---|
| 76 | if (u >= 256)
|
|---|
| 77 | return "???";
|
|---|
| 78 |
|
|---|
| 79 | char *pszRet;
|
|---|
| 80 | RTStrAPrintf(&pszRet, "%d way", u); /* intentional leak! */
|
|---|
| 81 | return pszRet;
|
|---|
| 82 | }
|
|---|
| 83 |
|
|---|
| 84 |
|
|---|
| 85 | const char *getL2CacheAss(unsigned u)
|
|---|
| 86 | {
|
|---|
| 87 | switch (u)
|
|---|
| 88 | {
|
|---|
| 89 | case 0: return "off ";
|
|---|
| 90 | case 1: return "direct";
|
|---|
| 91 | case 2: return "2 way ";
|
|---|
| 92 | case 3: return "res3 ";
|
|---|
| 93 | case 4: return "4 way ";
|
|---|
| 94 | case 5: return "res5 ";
|
|---|
| 95 | case 6: return "8 way ";
|
|---|
| 96 | case 7: return "res7 ";
|
|---|
| 97 | case 8: return "16 way";
|
|---|
| 98 | case 9: return "res9 ";
|
|---|
| 99 | case 10: return "res10 ";
|
|---|
| 100 | case 11: return "res11 ";
|
|---|
| 101 | case 12: return "res12 ";
|
|---|
| 102 | case 13: return "res13 ";
|
|---|
| 103 | case 14: return "res14 ";
|
|---|
| 104 | case 15: return "fully ";
|
|---|
| 105 | default:
|
|---|
| 106 | return "????";
|
|---|
| 107 | }
|
|---|
| 108 | }
|
|---|
| 109 |
|
|---|
| 110 |
|
|---|
| 111 | /**
|
|---|
| 112 | * Test and dump all possible info from the CPUID instruction.
|
|---|
| 113 | *
|
|---|
| 114 | * @remark Bits shared with the libc cpuid.c program. This all written by me, so no worries.
|
|---|
| 115 | * @todo transform the dumping into a generic runtime function. We'll need it for logging!
|
|---|
| 116 | */
|
|---|
| 117 | void tstASMCpuId(void)
|
|---|
| 118 | {
|
|---|
| 119 | unsigned iBit;
|
|---|
| 120 | struct
|
|---|
| 121 | {
|
|---|
| 122 | uint32_t uEBX, uEAX, uEDX, uECX;
|
|---|
| 123 | } s;
|
|---|
| 124 | if (!ASMHasCpuId())
|
|---|
| 125 | {
|
|---|
| 126 | RTPrintf("tstInlineAsm: warning! CPU doesn't support CPUID\n");
|
|---|
| 127 | return;
|
|---|
| 128 | }
|
|---|
| 129 |
|
|---|
| 130 | /*
|
|---|
| 131 | * Try the 0 function and use that for checking the ASMCpuId_* variants.
|
|---|
| 132 | */
|
|---|
| 133 | ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 134 |
|
|---|
| 135 | uint32_t u32 = ASMCpuId_ECX(0);
|
|---|
| 136 | CHECKVAL(u32, s.uECX, "%x");
|
|---|
| 137 |
|
|---|
| 138 | u32 = ASMCpuId_EDX(0);
|
|---|
| 139 | CHECKVAL(u32, s.uEDX, "%x");
|
|---|
| 140 |
|
|---|
| 141 | uint32_t uECX2 = s.uECX - 1;
|
|---|
| 142 | uint32_t uEDX2 = s.uEDX - 1;
|
|---|
| 143 | ASMCpuId_ECX_EDX(0, &uECX2, &uEDX2);
|
|---|
| 144 |
|
|---|
| 145 | CHECKVAL(uECX2, s.uECX, "%x");
|
|---|
| 146 | CHECKVAL(uEDX2, s.uEDX, "%x");
|
|---|
| 147 |
|
|---|
| 148 | /*
|
|---|
| 149 | * Done testing, dump the information.
|
|---|
| 150 | */
|
|---|
| 151 | RTPrintf("tstInlineAsm: CPUID Dump\n");
|
|---|
| 152 | ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 153 | const uint32_t cFunctions = s.uEAX;
|
|---|
| 154 |
|
|---|
| 155 | /* raw dump */
|
|---|
| 156 | RTPrintf("\n"
|
|---|
| 157 | " RAW Standard CPUIDs\n"
|
|---|
| 158 | "Function eax ebx ecx edx\n");
|
|---|
| 159 | for (unsigned iStd = 0; iStd <= cFunctions + 3; iStd++)
|
|---|
| 160 | {
|
|---|
| 161 | ASMCpuId(iStd, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 162 | RTPrintf("%08x %08x %08x %08x %08x%s\n",
|
|---|
| 163 | iStd, s.uEAX, s.uEBX, s.uECX, s.uEDX, iStd <= cFunctions ? "" : "*");
|
|---|
| 164 | }
|
|---|
| 165 |
|
|---|
| 166 | /*
|
|---|
| 167 | * Understandable output
|
|---|
| 168 | */
|
|---|
| 169 | ASMCpuId(0, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 170 | RTPrintf("Name: %.04s%.04s%.04s\n"
|
|---|
| 171 | "Support: 0-%u\n",
|
|---|
| 172 | &s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
|
|---|
| 173 | bool const fIntel = ASMIsIntelCpuEx(s.uEBX, s.uECX, s.uEDX);
|
|---|
| 174 |
|
|---|
| 175 | /*
|
|---|
| 176 | * Get Features.
|
|---|
| 177 | */
|
|---|
| 178 | if (cFunctions >= 1)
|
|---|
| 179 | {
|
|---|
| 180 | ASMCpuId(1, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 181 | RTPrintf("Family: %#x \tExtended: %#x \tEffective: %#x\n"
|
|---|
| 182 | "Model: %#x \tExtended: %#x \tEffective: %#x\n"
|
|---|
| 183 | "Stepping: %d\n"
|
|---|
| 184 | "APIC ID: %#04x\n"
|
|---|
| 185 | "Logical CPUs: %d\n"
|
|---|
| 186 | "CLFLUSH Size: %d\n"
|
|---|
| 187 | "Brand ID: %#04x\n",
|
|---|
| 188 | (s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
|
|---|
| 189 | (s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
|
|---|
| 190 | ASMGetCpuStepping(s.uEAX),
|
|---|
| 191 | (s.uEBX >> 24) & 0xff,
|
|---|
| 192 | (s.uEBX >> 16) & 0xff,
|
|---|
| 193 | (s.uEBX >> 8) & 0xff,
|
|---|
| 194 | (s.uEBX >> 0) & 0xff);
|
|---|
| 195 |
|
|---|
| 196 | RTPrintf("Features EDX: ");
|
|---|
| 197 | if (s.uEDX & RT_BIT(0)) RTPrintf(" FPU");
|
|---|
| 198 | if (s.uEDX & RT_BIT(1)) RTPrintf(" VME");
|
|---|
| 199 | if (s.uEDX & RT_BIT(2)) RTPrintf(" DE");
|
|---|
| 200 | if (s.uEDX & RT_BIT(3)) RTPrintf(" PSE");
|
|---|
| 201 | if (s.uEDX & RT_BIT(4)) RTPrintf(" TSC");
|
|---|
| 202 | if (s.uEDX & RT_BIT(5)) RTPrintf(" MSR");
|
|---|
| 203 | if (s.uEDX & RT_BIT(6)) RTPrintf(" PAE");
|
|---|
| 204 | if (s.uEDX & RT_BIT(7)) RTPrintf(" MCE");
|
|---|
| 205 | if (s.uEDX & RT_BIT(8)) RTPrintf(" CX8");
|
|---|
| 206 | if (s.uEDX & RT_BIT(9)) RTPrintf(" APIC");
|
|---|
| 207 | if (s.uEDX & RT_BIT(10)) RTPrintf(" 10");
|
|---|
| 208 | if (s.uEDX & RT_BIT(11)) RTPrintf(" SEP");
|
|---|
| 209 | if (s.uEDX & RT_BIT(12)) RTPrintf(" MTRR");
|
|---|
| 210 | if (s.uEDX & RT_BIT(13)) RTPrintf(" PGE");
|
|---|
| 211 | if (s.uEDX & RT_BIT(14)) RTPrintf(" MCA");
|
|---|
| 212 | if (s.uEDX & RT_BIT(15)) RTPrintf(" CMOV");
|
|---|
| 213 | if (s.uEDX & RT_BIT(16)) RTPrintf(" PAT");
|
|---|
| 214 | if (s.uEDX & RT_BIT(17)) RTPrintf(" PSE36");
|
|---|
| 215 | if (s.uEDX & RT_BIT(18)) RTPrintf(" PSN");
|
|---|
| 216 | if (s.uEDX & RT_BIT(19)) RTPrintf(" CLFSH");
|
|---|
| 217 | if (s.uEDX & RT_BIT(20)) RTPrintf(" 20");
|
|---|
| 218 | if (s.uEDX & RT_BIT(21)) RTPrintf(" DS");
|
|---|
| 219 | if (s.uEDX & RT_BIT(22)) RTPrintf(" ACPI");
|
|---|
| 220 | if (s.uEDX & RT_BIT(23)) RTPrintf(" MMX");
|
|---|
| 221 | if (s.uEDX & RT_BIT(24)) RTPrintf(" FXSR");
|
|---|
| 222 | if (s.uEDX & RT_BIT(25)) RTPrintf(" SSE");
|
|---|
| 223 | if (s.uEDX & RT_BIT(26)) RTPrintf(" SSE2");
|
|---|
| 224 | if (s.uEDX & RT_BIT(27)) RTPrintf(" SS");
|
|---|
| 225 | if (s.uEDX & RT_BIT(28)) RTPrintf(" HTT");
|
|---|
| 226 | if (s.uEDX & RT_BIT(29)) RTPrintf(" 29");
|
|---|
| 227 | if (s.uEDX & RT_BIT(30)) RTPrintf(" 30");
|
|---|
| 228 | if (s.uEDX & RT_BIT(31)) RTPrintf(" 31");
|
|---|
| 229 | RTPrintf("\n");
|
|---|
| 230 |
|
|---|
| 231 | /** @todo check intel docs. */
|
|---|
| 232 | RTPrintf("Features ECX: ");
|
|---|
| 233 | if (s.uECX & RT_BIT(0)) RTPrintf(" SSE3");
|
|---|
| 234 | for (iBit = 1; iBit < 13; iBit++)
|
|---|
| 235 | if (s.uECX & RT_BIT(iBit))
|
|---|
| 236 | RTPrintf(" %d", iBit);
|
|---|
| 237 | if (s.uECX & RT_BIT(13)) RTPrintf(" CX16");
|
|---|
| 238 | for (iBit = 14; iBit < 32; iBit++)
|
|---|
| 239 | if (s.uECX & RT_BIT(iBit))
|
|---|
| 240 | RTPrintf(" %d", iBit);
|
|---|
| 241 | RTPrintf("\n");
|
|---|
| 242 | }
|
|---|
| 243 |
|
|---|
| 244 | /*
|
|---|
| 245 | * Extended.
|
|---|
| 246 | * Implemented after AMD specs.
|
|---|
| 247 | */
|
|---|
| 248 | /** @todo check out the intel specs. */
|
|---|
| 249 | ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 250 | if (!s.uEAX && !s.uEBX && !s.uECX && !s.uEDX)
|
|---|
| 251 | {
|
|---|
| 252 | RTPrintf("No extended CPUID info? Check the manual on how to detect this...\n");
|
|---|
| 253 | return;
|
|---|
| 254 | }
|
|---|
| 255 | const uint32_t cExtFunctions = s.uEAX | 0x80000000;
|
|---|
| 256 |
|
|---|
| 257 | /* raw dump */
|
|---|
| 258 | RTPrintf("\n"
|
|---|
| 259 | " RAW Extended CPUIDs\n"
|
|---|
| 260 | "Function eax ebx ecx edx\n");
|
|---|
| 261 | for (unsigned iExt = 0x80000000; iExt <= cExtFunctions + 3; iExt++)
|
|---|
| 262 | {
|
|---|
| 263 | ASMCpuId(iExt, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 264 | RTPrintf("%08x %08x %08x %08x %08x%s\n",
|
|---|
| 265 | iExt, s.uEAX, s.uEBX, s.uECX, s.uEDX, iExt <= cExtFunctions ? "" : "*");
|
|---|
| 266 | }
|
|---|
| 267 |
|
|---|
| 268 | /*
|
|---|
| 269 | * Understandable output
|
|---|
| 270 | */
|
|---|
| 271 | ASMCpuId(0x80000000, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 272 | RTPrintf("Ext Name: %.4s%.4s%.4s\n"
|
|---|
| 273 | "Ext Supports: 0x80000000-%#010x\n",
|
|---|
| 274 | &s.uEBX, &s.uEDX, &s.uECX, s.uEAX);
|
|---|
| 275 |
|
|---|
| 276 | if (cExtFunctions >= 0x80000001)
|
|---|
| 277 | {
|
|---|
| 278 | ASMCpuId(0x80000001, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 279 | RTPrintf("Family: %#x \tExtended: %#x \tEffective: %#x\n"
|
|---|
| 280 | "Model: %#x \tExtended: %#x \tEffective: %#x\n"
|
|---|
| 281 | "Stepping: %d\n"
|
|---|
| 282 | "Brand ID: %#05x\n",
|
|---|
| 283 | (s.uEAX >> 8) & 0xf, (s.uEAX >> 20) & 0x7f, ASMGetCpuFamily(s.uEAX),
|
|---|
| 284 | (s.uEAX >> 4) & 0xf, (s.uEAX >> 16) & 0x0f, ASMGetCpuModel(s.uEAX, fIntel),
|
|---|
| 285 | ASMGetCpuStepping(s.uEAX),
|
|---|
| 286 | s.uEBX & 0xfff);
|
|---|
| 287 |
|
|---|
| 288 | RTPrintf("Features EDX: ");
|
|---|
| 289 | if (s.uEDX & RT_BIT(0)) RTPrintf(" FPU");
|
|---|
| 290 | if (s.uEDX & RT_BIT(1)) RTPrintf(" VME");
|
|---|
| 291 | if (s.uEDX & RT_BIT(2)) RTPrintf(" DE");
|
|---|
| 292 | if (s.uEDX & RT_BIT(3)) RTPrintf(" PSE");
|
|---|
| 293 | if (s.uEDX & RT_BIT(4)) RTPrintf(" TSC");
|
|---|
| 294 | if (s.uEDX & RT_BIT(5)) RTPrintf(" MSR");
|
|---|
| 295 | if (s.uEDX & RT_BIT(6)) RTPrintf(" PAE");
|
|---|
| 296 | if (s.uEDX & RT_BIT(7)) RTPrintf(" MCE");
|
|---|
| 297 | if (s.uEDX & RT_BIT(8)) RTPrintf(" CMPXCHG8B");
|
|---|
| 298 | if (s.uEDX & RT_BIT(9)) RTPrintf(" APIC");
|
|---|
| 299 | if (s.uEDX & RT_BIT(10)) RTPrintf(" 10");
|
|---|
| 300 | if (s.uEDX & RT_BIT(11)) RTPrintf(" SysCallSysRet");
|
|---|
| 301 | if (s.uEDX & RT_BIT(12)) RTPrintf(" MTRR");
|
|---|
| 302 | if (s.uEDX & RT_BIT(13)) RTPrintf(" PGE");
|
|---|
| 303 | if (s.uEDX & RT_BIT(14)) RTPrintf(" MCA");
|
|---|
| 304 | if (s.uEDX & RT_BIT(15)) RTPrintf(" CMOV");
|
|---|
| 305 | if (s.uEDX & RT_BIT(16)) RTPrintf(" PAT");
|
|---|
| 306 | if (s.uEDX & RT_BIT(17)) RTPrintf(" PSE36");
|
|---|
| 307 | if (s.uEDX & RT_BIT(18)) RTPrintf(" 18");
|
|---|
| 308 | if (s.uEDX & RT_BIT(19)) RTPrintf(" 19");
|
|---|
| 309 | if (s.uEDX & RT_BIT(20)) RTPrintf(" NX");
|
|---|
| 310 | if (s.uEDX & RT_BIT(21)) RTPrintf(" 21");
|
|---|
| 311 | if (s.uEDX & RT_BIT(22)) RTPrintf(" MmxExt");
|
|---|
| 312 | if (s.uEDX & RT_BIT(23)) RTPrintf(" MMX");
|
|---|
| 313 | if (s.uEDX & RT_BIT(24)) RTPrintf(" FXSR");
|
|---|
| 314 | if (s.uEDX & RT_BIT(25)) RTPrintf(" FastFXSR");
|
|---|
| 315 | if (s.uEDX & RT_BIT(26)) RTPrintf(" 26");
|
|---|
| 316 | if (s.uEDX & RT_BIT(27)) RTPrintf(" RDTSCP");
|
|---|
| 317 | if (s.uEDX & RT_BIT(28)) RTPrintf(" 28");
|
|---|
| 318 | if (s.uEDX & RT_BIT(29)) RTPrintf(" LongMode");
|
|---|
| 319 | if (s.uEDX & RT_BIT(30)) RTPrintf(" 3DNowExt");
|
|---|
| 320 | if (s.uEDX & RT_BIT(31)) RTPrintf(" 3DNow");
|
|---|
| 321 | RTPrintf("\n");
|
|---|
| 322 |
|
|---|
| 323 | RTPrintf("Features ECX: ");
|
|---|
| 324 | if (s.uECX & RT_BIT(0)) RTPrintf(" LahfSahf");
|
|---|
| 325 | if (s.uECX & RT_BIT(1)) RTPrintf(" CmpLegacy");
|
|---|
| 326 | if (s.uECX & RT_BIT(2)) RTPrintf(" SVM");
|
|---|
| 327 | if (s.uECX & RT_BIT(3)) RTPrintf(" 3");
|
|---|
| 328 | if (s.uECX & RT_BIT(4)) RTPrintf(" AltMovCr8");
|
|---|
| 329 | for (iBit = 5; iBit < 32; iBit++)
|
|---|
| 330 | if (s.uECX & RT_BIT(iBit))
|
|---|
| 331 | RTPrintf(" %d", iBit);
|
|---|
| 332 | RTPrintf("\n");
|
|---|
| 333 | }
|
|---|
| 334 |
|
|---|
| 335 | char szString[4*4*3+1] = {0};
|
|---|
| 336 | if (cExtFunctions >= 0x80000002)
|
|---|
| 337 | ASMCpuId(0x80000002, &szString[0 + 0], &szString[0 + 4], &szString[0 + 8], &szString[0 + 12]);
|
|---|
| 338 | if (cExtFunctions >= 0x80000003)
|
|---|
| 339 | ASMCpuId(0x80000003, &szString[16 + 0], &szString[16 + 4], &szString[16 + 8], &szString[16 + 12]);
|
|---|
| 340 | if (cExtFunctions >= 0x80000004)
|
|---|
| 341 | ASMCpuId(0x80000004, &szString[32 + 0], &szString[32 + 4], &szString[32 + 8], &szString[32 + 12]);
|
|---|
| 342 | if (cExtFunctions >= 0x80000002)
|
|---|
| 343 | RTPrintf("Full Name: %s\n", szString);
|
|---|
| 344 |
|
|---|
| 345 | if (cExtFunctions >= 0x80000005)
|
|---|
| 346 | {
|
|---|
| 347 | ASMCpuId(0x80000005, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 348 | RTPrintf("TLB 2/4M Instr/Uni: %s %3d entries\n"
|
|---|
| 349 | "TLB 2/4M Data: %s %3d entries\n",
|
|---|
| 350 | getCacheAss((s.uEAX >> 8) & 0xff), (s.uEAX >> 0) & 0xff,
|
|---|
| 351 | getCacheAss((s.uEAX >> 24) & 0xff), (s.uEAX >> 16) & 0xff);
|
|---|
| 352 | RTPrintf("TLB 4K Instr/Uni: %s %3d entries\n"
|
|---|
| 353 | "TLB 4K Data: %s %3d entries\n",
|
|---|
| 354 | getCacheAss((s.uEBX >> 8) & 0xff), (s.uEBX >> 0) & 0xff,
|
|---|
| 355 | getCacheAss((s.uEBX >> 24) & 0xff), (s.uEBX >> 16) & 0xff);
|
|---|
| 356 | RTPrintf("L1 Instr Cache Line Size: %d bytes\n"
|
|---|
| 357 | "L1 Instr Cache Lines Per Tag: %d\n"
|
|---|
| 358 | "L1 Instr Cache Associativity: %s\n"
|
|---|
| 359 | "L1 Instr Cache Size: %d KB\n",
|
|---|
| 360 | (s.uEDX >> 0) & 0xff,
|
|---|
| 361 | (s.uEDX >> 8) & 0xff,
|
|---|
| 362 | getCacheAss((s.uEDX >> 16) & 0xff),
|
|---|
| 363 | (s.uEDX >> 24) & 0xff);
|
|---|
| 364 | RTPrintf("L1 Data Cache Line Size: %d bytes\n"
|
|---|
| 365 | "L1 Data Cache Lines Per Tag: %d\n"
|
|---|
| 366 | "L1 Data Cache Associativity: %s\n"
|
|---|
| 367 | "L1 Data Cache Size: %d KB\n",
|
|---|
| 368 | (s.uECX >> 0) & 0xff,
|
|---|
| 369 | (s.uECX >> 8) & 0xff,
|
|---|
| 370 | getCacheAss((s.uECX >> 16) & 0xff),
|
|---|
| 371 | (s.uECX >> 24) & 0xff);
|
|---|
| 372 | }
|
|---|
| 373 |
|
|---|
| 374 | if (cExtFunctions >= 0x80000006)
|
|---|
| 375 | {
|
|---|
| 376 | ASMCpuId(0x80000006, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 377 | RTPrintf("L2 TLB 2/4M Instr/Uni: %s %4d entries\n"
|
|---|
| 378 | "L2 TLB 2/4M Data: %s %4d entries\n",
|
|---|
| 379 | getL2CacheAss((s.uEAX >> 12) & 0xf), (s.uEAX >> 0) & 0xfff,
|
|---|
| 380 | getL2CacheAss((s.uEAX >> 28) & 0xf), (s.uEAX >> 16) & 0xfff);
|
|---|
| 381 | RTPrintf("L2 TLB 4K Instr/Uni: %s %4d entries\n"
|
|---|
| 382 | "L2 TLB 4K Data: %s %4d entries\n",
|
|---|
| 383 | getL2CacheAss((s.uEBX >> 12) & 0xf), (s.uEBX >> 0) & 0xfff,
|
|---|
| 384 | getL2CacheAss((s.uEBX >> 28) & 0xf), (s.uEBX >> 16) & 0xfff);
|
|---|
| 385 | RTPrintf("L2 Cache Line Size: %d bytes\n"
|
|---|
| 386 | "L2 Cache Lines Per Tag: %d\n"
|
|---|
| 387 | "L2 Cache Associativity: %s\n"
|
|---|
| 388 | "L2 Cache Size: %d KB\n",
|
|---|
| 389 | (s.uEDX >> 0) & 0xff,
|
|---|
| 390 | (s.uEDX >> 8) & 0xf,
|
|---|
| 391 | getL2CacheAss((s.uEDX >> 12) & 0xf),
|
|---|
| 392 | (s.uEDX >> 16) & 0xffff);
|
|---|
| 393 | }
|
|---|
| 394 |
|
|---|
| 395 | if (cExtFunctions >= 0x80000007)
|
|---|
| 396 | {
|
|---|
| 397 | ASMCpuId(0x80000007, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 398 | RTPrintf("APM Features: ");
|
|---|
| 399 | if (s.uEDX & RT_BIT(0)) RTPrintf(" TS");
|
|---|
| 400 | if (s.uEDX & RT_BIT(1)) RTPrintf(" FID");
|
|---|
| 401 | if (s.uEDX & RT_BIT(2)) RTPrintf(" VID");
|
|---|
| 402 | if (s.uEDX & RT_BIT(3)) RTPrintf(" TTP");
|
|---|
| 403 | if (s.uEDX & RT_BIT(4)) RTPrintf(" TM");
|
|---|
| 404 | if (s.uEDX & RT_BIT(5)) RTPrintf(" STC");
|
|---|
| 405 | if (s.uEDX & RT_BIT(6)) RTPrintf(" 6");
|
|---|
| 406 | if (s.uEDX & RT_BIT(7)) RTPrintf(" 7");
|
|---|
| 407 | if (s.uEDX & RT_BIT(8)) RTPrintf(" TscInvariant");
|
|---|
| 408 | for (iBit = 9; iBit < 32; iBit++)
|
|---|
| 409 | if (s.uEDX & RT_BIT(iBit))
|
|---|
| 410 | RTPrintf(" %d", iBit);
|
|---|
| 411 | RTPrintf("\n");
|
|---|
| 412 | }
|
|---|
| 413 |
|
|---|
| 414 | if (cExtFunctions >= 0x80000008)
|
|---|
| 415 | {
|
|---|
| 416 | ASMCpuId(0x80000008, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 417 | RTPrintf("Physical Address Width: %d bits\n"
|
|---|
| 418 | "Virtual Address Width: %d bits\n",
|
|---|
| 419 | (s.uEAX >> 0) & 0xff,
|
|---|
| 420 | (s.uEAX >> 8) & 0xff);
|
|---|
| 421 | RTPrintf("Physical Core Count: %d\n",
|
|---|
| 422 | ((s.uECX >> 0) & 0xff) + 1);
|
|---|
| 423 | if ((s.uECX >> 12) & 0xf)
|
|---|
| 424 | RTPrintf("ApicIdCoreIdSize: %d bits\n", (s.uECX >> 12) & 0xf);
|
|---|
| 425 | }
|
|---|
| 426 |
|
|---|
| 427 | if (cExtFunctions >= 0x8000000a)
|
|---|
| 428 | {
|
|---|
| 429 | ASMCpuId(0x8000000a, &s.uEAX, &s.uEBX, &s.uECX, &s.uEDX);
|
|---|
| 430 | RTPrintf("SVM Revision: %d (%#x)\n"
|
|---|
| 431 | "Number of Address Space IDs: %d (%#x)\n",
|
|---|
| 432 | s.uEAX & 0xff, s.uEAX & 0xff,
|
|---|
| 433 | s.uEBX, s.uEBX);
|
|---|
| 434 | }
|
|---|
| 435 | }
|
|---|
| 436 | #endif /* !PIC || !X86 */
|
|---|
| 437 |
|
|---|
| 438 |
|
|---|
| 439 | static void tstASMAtomicXchgU8(void)
|
|---|
| 440 | {
|
|---|
| 441 | struct
|
|---|
| 442 | {
|
|---|
| 443 | uint8_t u8Dummy0;
|
|---|
| 444 | uint8_t u8;
|
|---|
| 445 | uint8_t u8Dummy1;
|
|---|
| 446 | } s;
|
|---|
| 447 |
|
|---|
| 448 | s.u8 = 0;
|
|---|
| 449 | s.u8Dummy0 = s.u8Dummy1 = 0x42;
|
|---|
| 450 | CHECKOP(ASMAtomicXchgU8(&s.u8, 1), 0, "%#x", uint8_t);
|
|---|
| 451 | CHECKVAL(s.u8, 1, "%#x");
|
|---|
| 452 |
|
|---|
| 453 | CHECKOP(ASMAtomicXchgU8(&s.u8, 0), 1, "%#x", uint8_t);
|
|---|
| 454 | CHECKVAL(s.u8, 0, "%#x");
|
|---|
| 455 |
|
|---|
| 456 | CHECKOP(ASMAtomicXchgU8(&s.u8, 0xff), 0, "%#x", uint8_t);
|
|---|
| 457 | CHECKVAL(s.u8, 0xff, "%#x");
|
|---|
| 458 |
|
|---|
| 459 | CHECKOP(ASMAtomicXchgU8(&s.u8, 0x87), 0xffff, "%#x", uint8_t);
|
|---|
| 460 | CHECKVAL(s.u8, 0x87, "%#x");
|
|---|
| 461 | CHECKVAL(s.u8Dummy0, 0x42, "%#x");
|
|---|
| 462 | CHECKVAL(s.u8Dummy1, 0x42, "%#x");
|
|---|
| 463 | }
|
|---|
| 464 |
|
|---|
| 465 |
|
|---|
| 466 | static void tstASMAtomicXchgU16(void)
|
|---|
| 467 | {
|
|---|
| 468 | struct
|
|---|
| 469 | {
|
|---|
| 470 | uint16_t u16Dummy0;
|
|---|
| 471 | uint16_t u16;
|
|---|
| 472 | uint16_t u16Dummy1;
|
|---|
| 473 | } s;
|
|---|
| 474 |
|
|---|
| 475 | s.u16 = 0;
|
|---|
| 476 | s.u16Dummy0 = s.u16Dummy1 = 0x1234;
|
|---|
| 477 | CHECKOP(ASMAtomicXchgU16(&s.u16, 1), 0, "%#x", uint16_t);
|
|---|
| 478 | CHECKVAL(s.u16, 1, "%#x");
|
|---|
| 479 |
|
|---|
| 480 | CHECKOP(ASMAtomicXchgU16(&s.u16, 0), 1, "%#x", uint16_t);
|
|---|
| 481 | CHECKVAL(s.u16, 0, "%#x");
|
|---|
| 482 |
|
|---|
| 483 | CHECKOP(ASMAtomicXchgU16(&s.u16, 0xffff), 0, "%#x", uint16_t);
|
|---|
| 484 | CHECKVAL(s.u16, 0xffff, "%#x");
|
|---|
| 485 |
|
|---|
| 486 | CHECKOP(ASMAtomicXchgU16(&s.u16, 0x8765), 0xffff, "%#x", uint16_t);
|
|---|
| 487 | CHECKVAL(s.u16, 0x8765, "%#x");
|
|---|
| 488 | CHECKVAL(s.u16Dummy0, 0x1234, "%#x");
|
|---|
| 489 | CHECKVAL(s.u16Dummy1, 0x1234, "%#x");
|
|---|
| 490 | }
|
|---|
| 491 |
|
|---|
| 492 |
|
|---|
| 493 | static void tstASMAtomicXchgU32(void)
|
|---|
| 494 | {
|
|---|
| 495 | struct
|
|---|
| 496 | {
|
|---|
| 497 | uint32_t u32Dummy0;
|
|---|
| 498 | uint32_t u32;
|
|---|
| 499 | uint32_t u32Dummy1;
|
|---|
| 500 | } s;
|
|---|
| 501 |
|
|---|
| 502 | s.u32 = 0;
|
|---|
| 503 | s.u32Dummy0 = s.u32Dummy1 = 0x11223344;
|
|---|
| 504 |
|
|---|
| 505 | CHECKOP(ASMAtomicXchgU32(&s.u32, 1), 0, "%#x", uint32_t);
|
|---|
| 506 | CHECKVAL(s.u32, 1, "%#x");
|
|---|
| 507 |
|
|---|
| 508 | CHECKOP(ASMAtomicXchgU32(&s.u32, 0), 1, "%#x", uint32_t);
|
|---|
| 509 | CHECKVAL(s.u32, 0, "%#x");
|
|---|
| 510 |
|
|---|
| 511 | CHECKOP(ASMAtomicXchgU32(&s.u32, ~0U), 0, "%#x", uint32_t);
|
|---|
| 512 | CHECKVAL(s.u32, ~0U, "%#x");
|
|---|
| 513 |
|
|---|
| 514 | CHECKOP(ASMAtomicXchgU32(&s.u32, 0x87654321), ~0U, "%#x", uint32_t);
|
|---|
| 515 | CHECKVAL(s.u32, 0x87654321, "%#x");
|
|---|
| 516 |
|
|---|
| 517 | CHECKVAL(s.u32Dummy0, 0x11223344, "%#x");
|
|---|
| 518 | CHECKVAL(s.u32Dummy1, 0x11223344, "%#x");
|
|---|
| 519 | }
|
|---|
| 520 |
|
|---|
| 521 |
|
|---|
| 522 | static void tstASMAtomicXchgU64(void)
|
|---|
| 523 | {
|
|---|
| 524 | struct
|
|---|
| 525 | {
|
|---|
| 526 | uint64_t u64Dummy0;
|
|---|
| 527 | uint64_t u64;
|
|---|
| 528 | uint64_t u64Dummy1;
|
|---|
| 529 | } s;
|
|---|
| 530 |
|
|---|
| 531 | s.u64 = 0;
|
|---|
| 532 | s.u64Dummy0 = s.u64Dummy1 = 0x1122334455667788ULL;
|
|---|
| 533 |
|
|---|
| 534 | CHECKOP(ASMAtomicXchgU64(&s.u64, 1), 0ULL, "%#llx", uint64_t);
|
|---|
| 535 | CHECKVAL(s.u64, 1ULL, "%#llx");
|
|---|
| 536 |
|
|---|
| 537 | CHECKOP(ASMAtomicXchgU64(&s.u64, 0), 1ULL, "%#llx", uint64_t);
|
|---|
| 538 | CHECKVAL(s.u64, 0ULL, "%#llx");
|
|---|
| 539 |
|
|---|
| 540 | CHECKOP(ASMAtomicXchgU64(&s.u64, ~0ULL), 0ULL, "%#llx", uint64_t);
|
|---|
| 541 | CHECKVAL(s.u64, ~0ULL, "%#llx");
|
|---|
| 542 |
|
|---|
| 543 | CHECKOP(ASMAtomicXchgU64(&s.u64, 0xfedcba0987654321ULL), ~0ULL, "%#llx", uint64_t);
|
|---|
| 544 | CHECKVAL(s.u64, 0xfedcba0987654321ULL, "%#llx");
|
|---|
| 545 |
|
|---|
| 546 | CHECKVAL(s.u64Dummy0, 0x1122334455667788ULL, "%#llx");
|
|---|
| 547 | CHECKVAL(s.u64Dummy1, 0x1122334455667788ULL, "%#llx");
|
|---|
| 548 | }
|
|---|
| 549 |
|
|---|
| 550 |
|
|---|
| 551 | static void tstASMAtomicXchgPtr(void)
|
|---|
| 552 | {
|
|---|
| 553 | void *pv = NULL;
|
|---|
| 554 |
|
|---|
| 555 | CHECKOP(ASMAtomicXchgPtr(&pv, (void *)(~(uintptr_t)0)), NULL, "%p", void *);
|
|---|
| 556 | CHECKVAL(pv, (void *)(~(uintptr_t)0), "%p");
|
|---|
| 557 |
|
|---|
| 558 | CHECKOP(ASMAtomicXchgPtr(&pv, (void *)0x87654321), (void *)(~(uintptr_t)0), "%p", void *);
|
|---|
| 559 | CHECKVAL(pv, (void *)0x87654321, "%p");
|
|---|
| 560 |
|
|---|
| 561 | CHECKOP(ASMAtomicXchgPtr(&pv, NULL), (void *)0x87654321, "%p", void *);
|
|---|
| 562 | CHECKVAL(pv, NULL, "%p");
|
|---|
| 563 | }
|
|---|
| 564 |
|
|---|
| 565 |
|
|---|
| 566 | static void tstASMAtomicCmpXchgU8(void)
|
|---|
| 567 | {
|
|---|
| 568 | struct
|
|---|
| 569 | {
|
|---|
| 570 | uint8_t u8Before;
|
|---|
| 571 | uint8_t u8;
|
|---|
| 572 | uint8_t u8After;
|
|---|
| 573 | } u = { 0xcc, 0xff, 0xaa };
|
|---|
| 574 |
|
|---|
| 575 | CHECKOP(ASMAtomicCmpXchgU8(&u.u8, 0, 0), false, "%d", bool);
|
|---|
| 576 | CHECKVAL(u.u8, 0xff, "%x"); CHECKVAL(u.u8Before, 0xcc, "%x"); CHECKVAL(u.u8After, 0xaa, "%x");
|
|---|
| 577 |
|
|---|
| 578 | CHECKOP(ASMAtomicCmpXchgU8(&u.u8, 0, 0xff), true, "%d", bool);
|
|---|
| 579 | CHECKVAL(u.u8, 0, "%x"); CHECKVAL(u.u8Before, 0xcc, "%x"); CHECKVAL(u.u8After, 0xaa, "%x");
|
|---|
| 580 |
|
|---|
| 581 | CHECKOP(ASMAtomicCmpXchgU8(&u.u8, 0x79, 0xff), false, "%d", bool);
|
|---|
| 582 | CHECKVAL(u.u8, 0, "%x"); CHECKVAL(u.u8Before, 0xcc, "%x"); CHECKVAL(u.u8After, 0xaa, "%x");
|
|---|
| 583 |
|
|---|
| 584 | CHECKOP(ASMAtomicCmpXchgU8(&u.u8, 0x97, 0), true, "%d", bool);
|
|---|
| 585 | CHECKVAL(u.u8, 0x97, "%x"); CHECKVAL(u.u8Before, 0xcc, "%x"); CHECKVAL(u.u8After, 0xaa, "%x");
|
|---|
| 586 | }
|
|---|
| 587 |
|
|---|
| 588 | static void tstASMAtomicCmpXchgU32(void)
|
|---|
| 589 | {
|
|---|
| 590 | uint32_t u32 = 0xffffffff;
|
|---|
| 591 |
|
|---|
| 592 | CHECKOP(ASMAtomicCmpXchgU32(&u32, 0, 0), false, "%d", bool);
|
|---|
| 593 | CHECKVAL(u32, 0xffffffff, "%x");
|
|---|
| 594 |
|
|---|
| 595 | CHECKOP(ASMAtomicCmpXchgU32(&u32, 0, 0xffffffff), true, "%d", bool);
|
|---|
| 596 | CHECKVAL(u32, 0, "%x");
|
|---|
| 597 |
|
|---|
| 598 | CHECKOP(ASMAtomicCmpXchgU32(&u32, 0x8008efd, 0xffffffff), false, "%d", bool);
|
|---|
| 599 | CHECKVAL(u32, 0, "%x");
|
|---|
| 600 |
|
|---|
| 601 | CHECKOP(ASMAtomicCmpXchgU32(&u32, 0x8008efd, 0), true, "%d", bool);
|
|---|
| 602 | CHECKVAL(u32, 0x8008efd, "%x");
|
|---|
| 603 | }
|
|---|
| 604 |
|
|---|
| 605 |
|
|---|
| 606 | static void tstASMAtomicCmpXchgU64(void)
|
|---|
| 607 | {
|
|---|
| 608 | uint64_t u64 = 0xffffffffffffffULL;
|
|---|
| 609 |
|
|---|
| 610 | CHECKOP(ASMAtomicCmpXchgU64(&u64, 0, 0), false, "%d", bool);
|
|---|
| 611 | CHECKVAL(u64, 0xffffffffffffffULL, "%#llx");
|
|---|
| 612 |
|
|---|
| 613 | CHECKOP(ASMAtomicCmpXchgU64(&u64, 0, 0xffffffffffffffULL), true, "%d", bool);
|
|---|
| 614 | CHECKVAL(u64, 0, "%x");
|
|---|
| 615 |
|
|---|
| 616 | CHECKOP(ASMAtomicCmpXchgU64(&u64, 0x80040008008efdULL, 0xffffffff), false, "%d", bool);
|
|---|
| 617 | CHECKVAL(u64, 0, "%x");
|
|---|
| 618 |
|
|---|
| 619 | CHECKOP(ASMAtomicCmpXchgU64(&u64, 0x80040008008efdULL, 0xffffffff00000000ULL), false, "%d", bool);
|
|---|
| 620 | CHECKVAL(u64, 0, "%x");
|
|---|
| 621 |
|
|---|
| 622 | CHECKOP(ASMAtomicCmpXchgU64(&u64, 0x80040008008efdULL, 0), true, "%d", bool);
|
|---|
| 623 | CHECKVAL(u64, 0x80040008008efdULL, "%#llx");
|
|---|
| 624 | }
|
|---|
| 625 |
|
|---|
| 626 |
|
|---|
| 627 | static void tstASMAtomicCmpXchgExU32(void)
|
|---|
| 628 | {
|
|---|
| 629 | uint32_t u32 = 0xffffffff;
|
|---|
| 630 | uint32_t u32Old = 0x80005111;
|
|---|
| 631 |
|
|---|
| 632 | CHECKOP(ASMAtomicCmpXchgExU32(&u32, 0, 0, &u32Old), false, "%d", bool);
|
|---|
| 633 | CHECKVAL(u32, 0xffffffff, "%x");
|
|---|
| 634 | CHECKVAL(u32Old, 0xffffffff, "%x");
|
|---|
| 635 |
|
|---|
| 636 | CHECKOP(ASMAtomicCmpXchgExU32(&u32, 0, 0xffffffff, &u32Old), true, "%d", bool);
|
|---|
| 637 | CHECKVAL(u32, 0, "%x");
|
|---|
| 638 | CHECKVAL(u32Old, 0xffffffff, "%x");
|
|---|
| 639 |
|
|---|
| 640 | CHECKOP(ASMAtomicCmpXchgExU32(&u32, 0x8008efd, 0xffffffff, &u32Old), false, "%d", bool);
|
|---|
| 641 | CHECKVAL(u32, 0, "%x");
|
|---|
| 642 | CHECKVAL(u32Old, 0, "%x");
|
|---|
| 643 |
|
|---|
| 644 | CHECKOP(ASMAtomicCmpXchgExU32(&u32, 0x8008efd, 0, &u32Old), true, "%d", bool);
|
|---|
| 645 | CHECKVAL(u32, 0x8008efd, "%x");
|
|---|
| 646 | CHECKVAL(u32Old, 0, "%x");
|
|---|
| 647 |
|
|---|
| 648 | CHECKOP(ASMAtomicCmpXchgExU32(&u32, 0, 0x8008efd, &u32Old), true, "%d", bool);
|
|---|
| 649 | CHECKVAL(u32, 0, "%x");
|
|---|
| 650 | CHECKVAL(u32Old, 0x8008efd, "%x");
|
|---|
| 651 | }
|
|---|
| 652 |
|
|---|
| 653 |
|
|---|
| 654 | static void tstASMAtomicCmpXchgExU64(void)
|
|---|
| 655 | {
|
|---|
| 656 | uint64_t u64 = 0xffffffffffffffffULL;
|
|---|
| 657 | uint64_t u64Old = 0x8000000051111111ULL;
|
|---|
| 658 |
|
|---|
| 659 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0, 0, &u64Old), false, "%d", bool);
|
|---|
| 660 | CHECKVAL(u64, 0xffffffffffffffffULL, "%llx");
|
|---|
| 661 | CHECKVAL(u64Old, 0xffffffffffffffffULL, "%llx");
|
|---|
| 662 |
|
|---|
| 663 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0, 0xffffffffffffffffULL, &u64Old), true, "%d", bool);
|
|---|
| 664 | CHECKVAL(u64, 0ULL, "%llx");
|
|---|
| 665 | CHECKVAL(u64Old, 0xffffffffffffffffULL, "%llx");
|
|---|
| 666 |
|
|---|
| 667 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0x80040008008efdULL, 0xffffffff, &u64Old), false, "%d", bool);
|
|---|
| 668 | CHECKVAL(u64, 0ULL, "%llx");
|
|---|
| 669 | CHECKVAL(u64Old, 0ULL, "%llx");
|
|---|
| 670 |
|
|---|
| 671 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0x80040008008efdULL, 0xffffffff00000000ULL, &u64Old), false, "%d", bool);
|
|---|
| 672 | CHECKVAL(u64, 0ULL, "%llx");
|
|---|
| 673 | CHECKVAL(u64Old, 0ULL, "%llx");
|
|---|
| 674 |
|
|---|
| 675 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0x80040008008efdULL, 0, &u64Old), true, "%d", bool);
|
|---|
| 676 | CHECKVAL(u64, 0x80040008008efdULL, "%llx");
|
|---|
| 677 | CHECKVAL(u64Old, 0ULL, "%llx");
|
|---|
| 678 |
|
|---|
| 679 | CHECKOP(ASMAtomicCmpXchgExU64(&u64, 0, 0x80040008008efdULL, &u64Old), true, "%d", bool);
|
|---|
| 680 | CHECKVAL(u64, 0ULL, "%llx");
|
|---|
| 681 | CHECKVAL(u64Old, 0x80040008008efdULL, "%llx");
|
|---|
| 682 | }
|
|---|
| 683 |
|
|---|
| 684 |
|
|---|
| 685 | static void tstASMAtomicReadU64(void)
|
|---|
| 686 | {
|
|---|
| 687 | uint64_t u64 = 0;
|
|---|
| 688 |
|
|---|
| 689 | CHECKOP(ASMAtomicReadU64(&u64), 0ULL, "%#llx", uint64_t);
|
|---|
| 690 | CHECKVAL(u64, 0ULL, "%#llx");
|
|---|
| 691 |
|
|---|
| 692 | u64 = ~0ULL;
|
|---|
| 693 | CHECKOP(ASMAtomicReadU64(&u64), ~0ULL, "%#llx", uint64_t);
|
|---|
| 694 | CHECKVAL(u64, ~0ULL, "%#llx");
|
|---|
| 695 |
|
|---|
| 696 | u64 = 0xfedcba0987654321ULL;
|
|---|
| 697 | CHECKOP(ASMAtomicReadU64(&u64), 0xfedcba0987654321ULL, "%#llx", uint64_t);
|
|---|
| 698 | CHECKVAL(u64, 0xfedcba0987654321ULL, "%#llx");
|
|---|
| 699 | }
|
|---|
| 700 |
|
|---|
| 701 |
|
|---|
| 702 | static void tstASMAtomicAddS32(void)
|
|---|
| 703 | {
|
|---|
| 704 | int32_t i32Rc;
|
|---|
| 705 | int32_t i32 = 10;
|
|---|
| 706 | #define MYCHECK(op, rc, val) \
|
|---|
| 707 | do { \
|
|---|
| 708 | i32Rc = op; \
|
|---|
| 709 | if (i32Rc != (rc)) \
|
|---|
| 710 | { \
|
|---|
| 711 | RTPrintf("%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
|
|---|
| 712 | RTTestIErrorInc(); \
|
|---|
| 713 | } \
|
|---|
| 714 | if (i32 != (val)) \
|
|---|
| 715 | { \
|
|---|
| 716 | RTPrintf("%s, %d: FAILURE: %s => i32=%d expected %d\n", __FUNCTION__, __LINE__, #op, i32, val); \
|
|---|
| 717 | RTTestIErrorInc(); \
|
|---|
| 718 | } \
|
|---|
| 719 | } while (0)
|
|---|
| 720 | MYCHECK(ASMAtomicAddS32(&i32, 1), 10, 11);
|
|---|
| 721 | MYCHECK(ASMAtomicAddS32(&i32, -2), 11, 9);
|
|---|
| 722 | MYCHECK(ASMAtomicAddS32(&i32, -9), 9, 0);
|
|---|
| 723 | MYCHECK(ASMAtomicAddS32(&i32, -0x7fffffff), 0, -0x7fffffff);
|
|---|
| 724 | MYCHECK(ASMAtomicAddS32(&i32, 0), -0x7fffffff, -0x7fffffff);
|
|---|
| 725 | MYCHECK(ASMAtomicAddS32(&i32, 0x7fffffff), -0x7fffffff, 0);
|
|---|
| 726 | MYCHECK(ASMAtomicAddS32(&i32, 0), 0, 0);
|
|---|
| 727 | #undef MYCHECK
|
|---|
| 728 | }
|
|---|
| 729 |
|
|---|
| 730 |
|
|---|
| 731 | static void tstASMAtomicDecIncS32(void)
|
|---|
| 732 | {
|
|---|
| 733 | int32_t i32Rc;
|
|---|
| 734 | int32_t i32 = 10;
|
|---|
| 735 | #define MYCHECK(op, rc) \
|
|---|
| 736 | do { \
|
|---|
| 737 | i32Rc = op; \
|
|---|
| 738 | if (i32Rc != (rc)) \
|
|---|
| 739 | { \
|
|---|
| 740 | RTPrintf("%s, %d: FAILURE: %s -> %d expected %d\n", __FUNCTION__, __LINE__, #op, i32Rc, rc); \
|
|---|
| 741 | RTTestIErrorInc(); \
|
|---|
| 742 | } \
|
|---|
| 743 | if (i32 != (rc)) \
|
|---|
| 744 | { \
|
|---|
| 745 | RTPrintf("%s, %d: FAILURE: %s => i32=%d expected %d\n", __FUNCTION__, __LINE__, #op, i32, rc); \
|
|---|
| 746 | RTTestIErrorInc(); \
|
|---|
| 747 | } \
|
|---|
| 748 | } while (0)
|
|---|
| 749 | MYCHECK(ASMAtomicDecS32(&i32), 9);
|
|---|
| 750 | MYCHECK(ASMAtomicDecS32(&i32), 8);
|
|---|
| 751 | MYCHECK(ASMAtomicDecS32(&i32), 7);
|
|---|
| 752 | MYCHECK(ASMAtomicDecS32(&i32), 6);
|
|---|
| 753 | MYCHECK(ASMAtomicDecS32(&i32), 5);
|
|---|
| 754 | MYCHECK(ASMAtomicDecS32(&i32), 4);
|
|---|
| 755 | MYCHECK(ASMAtomicDecS32(&i32), 3);
|
|---|
| 756 | MYCHECK(ASMAtomicDecS32(&i32), 2);
|
|---|
| 757 | MYCHECK(ASMAtomicDecS32(&i32), 1);
|
|---|
| 758 | MYCHECK(ASMAtomicDecS32(&i32), 0);
|
|---|
| 759 | MYCHECK(ASMAtomicDecS32(&i32), -1);
|
|---|
| 760 | MYCHECK(ASMAtomicDecS32(&i32), -2);
|
|---|
| 761 | MYCHECK(ASMAtomicIncS32(&i32), -1);
|
|---|
| 762 | MYCHECK(ASMAtomicIncS32(&i32), 0);
|
|---|
| 763 | MYCHECK(ASMAtomicIncS32(&i32), 1);
|
|---|
| 764 | MYCHECK(ASMAtomicIncS32(&i32), 2);
|
|---|
| 765 | MYCHECK(ASMAtomicIncS32(&i32), 3);
|
|---|
| 766 | MYCHECK(ASMAtomicDecS32(&i32), 2);
|
|---|
| 767 | MYCHECK(ASMAtomicIncS32(&i32), 3);
|
|---|
| 768 | MYCHECK(ASMAtomicDecS32(&i32), 2);
|
|---|
| 769 | MYCHECK(ASMAtomicIncS32(&i32), 3);
|
|---|
| 770 | #undef MYCHECK
|
|---|
| 771 | }
|
|---|
| 772 |
|
|---|
| 773 |
|
|---|
| 774 | static void tstASMAtomicAndOrU32(void)
|
|---|
| 775 | {
|
|---|
| 776 | uint32_t u32 = 0xffffffff;
|
|---|
| 777 |
|
|---|
| 778 | ASMAtomicOrU32(&u32, 0xffffffff);
|
|---|
| 779 | CHECKVAL(u32, 0xffffffff, "%x");
|
|---|
| 780 |
|
|---|
| 781 | ASMAtomicAndU32(&u32, 0xffffffff);
|
|---|
| 782 | CHECKVAL(u32, 0xffffffff, "%x");
|
|---|
| 783 |
|
|---|
| 784 | ASMAtomicAndU32(&u32, 0x8f8f8f8f);
|
|---|
| 785 | CHECKVAL(u32, 0x8f8f8f8f, "%x");
|
|---|
| 786 |
|
|---|
| 787 | ASMAtomicOrU32(&u32, 0x70707070);
|
|---|
| 788 | CHECKVAL(u32, 0xffffffff, "%x");
|
|---|
| 789 |
|
|---|
| 790 | ASMAtomicAndU32(&u32, 1);
|
|---|
| 791 | CHECKVAL(u32, 1, "%x");
|
|---|
| 792 |
|
|---|
| 793 | ASMAtomicOrU32(&u32, 0x80000000);
|
|---|
| 794 | CHECKVAL(u32, 0x80000001, "%x");
|
|---|
| 795 |
|
|---|
| 796 | ASMAtomicAndU32(&u32, 0x80000000);
|
|---|
| 797 | CHECKVAL(u32, 0x80000000, "%x");
|
|---|
| 798 |
|
|---|
| 799 | ASMAtomicAndU32(&u32, 0);
|
|---|
| 800 | CHECKVAL(u32, 0, "%x");
|
|---|
| 801 |
|
|---|
| 802 | ASMAtomicOrU32(&u32, 0x42424242);
|
|---|
| 803 | CHECKVAL(u32, 0x42424242, "%x");
|
|---|
| 804 | }
|
|---|
| 805 |
|
|---|
| 806 |
|
|---|
| 807 | void tstASMMemZeroPage(void)
|
|---|
| 808 | {
|
|---|
| 809 | struct
|
|---|
| 810 | {
|
|---|
| 811 | uint64_t u64Magic1;
|
|---|
| 812 | uint8_t abPage[PAGE_SIZE];
|
|---|
| 813 | uint64_t u64Magic2;
|
|---|
| 814 | } Buf1, Buf2, Buf3;
|
|---|
| 815 |
|
|---|
| 816 | Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 817 | memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
|
|---|
| 818 | Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 819 | Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 820 | memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
|
|---|
| 821 | Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 822 | Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 823 | memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
|
|---|
| 824 | Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 825 | ASMMemZeroPage(Buf1.abPage);
|
|---|
| 826 | ASMMemZeroPage(Buf2.abPage);
|
|---|
| 827 | ASMMemZeroPage(Buf3.abPage);
|
|---|
| 828 | if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 829 | || Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 830 | || Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 831 | || Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 832 | || Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 833 | || Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
|
|---|
| 834 | {
|
|---|
| 835 | RTPrintf("tstInlineAsm: ASMMemZeroPage violated one/both magic(s)!\n");
|
|---|
| 836 | RTTestIErrorInc();
|
|---|
| 837 | }
|
|---|
| 838 | for (unsigned i = 0; i < sizeof(Buf1.abPage); i++)
|
|---|
| 839 | if (Buf1.abPage[i])
|
|---|
| 840 | {
|
|---|
| 841 | RTPrintf("tstInlineAsm: ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
|
|---|
| 842 | RTTestIErrorInc();
|
|---|
| 843 | }
|
|---|
| 844 | for (unsigned i = 0; i < sizeof(Buf2.abPage); i++)
|
|---|
| 845 | if (Buf2.abPage[i])
|
|---|
| 846 | {
|
|---|
| 847 | RTPrintf("tstInlineAsm: ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
|
|---|
| 848 | RTTestIErrorInc();
|
|---|
| 849 | }
|
|---|
| 850 | for (unsigned i = 0; i < sizeof(Buf3.abPage); i++)
|
|---|
| 851 | if (Buf3.abPage[i])
|
|---|
| 852 | {
|
|---|
| 853 | RTPrintf("tstInlineAsm: ASMMemZeroPage didn't clear byte at offset %#x!\n", i);
|
|---|
| 854 | RTTestIErrorInc();
|
|---|
| 855 | }
|
|---|
| 856 | }
|
|---|
| 857 |
|
|---|
| 858 |
|
|---|
| 859 | void tstASMMemIsZeroPage(RTTEST hTest)
|
|---|
| 860 | {
|
|---|
| 861 | RTTestSub(hTest, "ASMMemIsZeroPage");
|
|---|
| 862 |
|
|---|
| 863 | void *pvPage1 = RTTestGuardedAllocHead(hTest, PAGE_SIZE);
|
|---|
| 864 | void *pvPage2 = RTTestGuardedAllocTail(hTest, PAGE_SIZE);
|
|---|
| 865 | RTTESTI_CHECK_RETV(pvPage1 && pvPage2);
|
|---|
| 866 |
|
|---|
| 867 | memset(pvPage1, 0, PAGE_SIZE);
|
|---|
| 868 | memset(pvPage2, 0, PAGE_SIZE);
|
|---|
| 869 | RTTESTI_CHECK(ASMMemIsZeroPage(pvPage1));
|
|---|
| 870 | RTTESTI_CHECK(ASMMemIsZeroPage(pvPage2));
|
|---|
| 871 |
|
|---|
| 872 | memset(pvPage1, 0xff, PAGE_SIZE);
|
|---|
| 873 | memset(pvPage2, 0xff, PAGE_SIZE);
|
|---|
| 874 | RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
|
|---|
| 875 | RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
|
|---|
| 876 |
|
|---|
| 877 | memset(pvPage1, 0, PAGE_SIZE);
|
|---|
| 878 | memset(pvPage2, 0, PAGE_SIZE);
|
|---|
| 879 | for (unsigned off = 0; off < PAGE_SIZE; off++)
|
|---|
| 880 | {
|
|---|
| 881 | ((uint8_t *)pvPage1)[off] = 1;
|
|---|
| 882 | RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage1));
|
|---|
| 883 | ((uint8_t *)pvPage1)[off] = 0;
|
|---|
| 884 |
|
|---|
| 885 | ((uint8_t *)pvPage2)[off] = 0x80;
|
|---|
| 886 | RTTESTI_CHECK(!ASMMemIsZeroPage(pvPage2));
|
|---|
| 887 | ((uint8_t *)pvPage2)[off] = 0;
|
|---|
| 888 | }
|
|---|
| 889 |
|
|---|
| 890 | RTTestSubDone(hTest);
|
|---|
| 891 | }
|
|---|
| 892 |
|
|---|
| 893 |
|
|---|
| 894 | void tstASMMemZero32(void)
|
|---|
| 895 | {
|
|---|
| 896 | struct
|
|---|
| 897 | {
|
|---|
| 898 | uint64_t u64Magic1;
|
|---|
| 899 | uint8_t abPage[PAGE_SIZE - 32];
|
|---|
| 900 | uint64_t u64Magic2;
|
|---|
| 901 | } Buf1, Buf2, Buf3;
|
|---|
| 902 |
|
|---|
| 903 | Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 904 | memset(Buf1.abPage, 0x55, sizeof(Buf1.abPage));
|
|---|
| 905 | Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 906 | Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 907 | memset(Buf2.abPage, 0x77, sizeof(Buf2.abPage));
|
|---|
| 908 | Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 909 | Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 910 | memset(Buf3.abPage, 0x99, sizeof(Buf3.abPage));
|
|---|
| 911 | Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 912 | ASMMemZero32(Buf1.abPage, sizeof(Buf1.abPage));
|
|---|
| 913 | ASMMemZero32(Buf2.abPage, sizeof(Buf2.abPage));
|
|---|
| 914 | ASMMemZero32(Buf3.abPage, sizeof(Buf3.abPage));
|
|---|
| 915 | if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 916 | || Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 917 | || Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 918 | || Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 919 | || Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 920 | || Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
|
|---|
| 921 | {
|
|---|
| 922 | RTPrintf("tstInlineAsm: ASMMemZero32 violated one/both magic(s)!\n");
|
|---|
| 923 | RTTestIErrorInc();
|
|---|
| 924 | }
|
|---|
| 925 | for (unsigned i = 0; i < RT_ELEMENTS(Buf1.abPage); i++)
|
|---|
| 926 | if (Buf1.abPage[i])
|
|---|
| 927 | {
|
|---|
| 928 | RTPrintf("tstInlineAsm: ASMMemZero32 didn't clear byte at offset %#x!\n", i);
|
|---|
| 929 | RTTestIErrorInc();
|
|---|
| 930 | }
|
|---|
| 931 | for (unsigned i = 0; i < RT_ELEMENTS(Buf2.abPage); i++)
|
|---|
| 932 | if (Buf2.abPage[i])
|
|---|
| 933 | {
|
|---|
| 934 | RTPrintf("tstInlineAsm: ASMMemZero32 didn't clear byte at offset %#x!\n", i);
|
|---|
| 935 | RTTestIErrorInc();
|
|---|
| 936 | }
|
|---|
| 937 | for (unsigned i = 0; i < RT_ELEMENTS(Buf3.abPage); i++)
|
|---|
| 938 | if (Buf3.abPage[i])
|
|---|
| 939 | {
|
|---|
| 940 | RTPrintf("tstInlineAsm: ASMMemZero32 didn't clear byte at offset %#x!\n", i);
|
|---|
| 941 | RTTestIErrorInc();
|
|---|
| 942 | }
|
|---|
| 943 | }
|
|---|
| 944 |
|
|---|
| 945 |
|
|---|
| 946 | void tstASMMemFill32(void)
|
|---|
| 947 | {
|
|---|
| 948 | struct
|
|---|
| 949 | {
|
|---|
| 950 | uint64_t u64Magic1;
|
|---|
| 951 | uint32_t au32Page[PAGE_SIZE / 4];
|
|---|
| 952 | uint64_t u64Magic2;
|
|---|
| 953 | } Buf1;
|
|---|
| 954 | struct
|
|---|
| 955 | {
|
|---|
| 956 | uint64_t u64Magic1;
|
|---|
| 957 | uint32_t au32Page[(PAGE_SIZE / 4) - 3];
|
|---|
| 958 | uint64_t u64Magic2;
|
|---|
| 959 | } Buf2;
|
|---|
| 960 | struct
|
|---|
| 961 | {
|
|---|
| 962 | uint64_t u64Magic1;
|
|---|
| 963 | uint32_t au32Page[(PAGE_SIZE / 4) - 1];
|
|---|
| 964 | uint64_t u64Magic2;
|
|---|
| 965 | } Buf3;
|
|---|
| 966 |
|
|---|
| 967 | Buf1.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 968 | memset(Buf1.au32Page, 0x55, sizeof(Buf1.au32Page));
|
|---|
| 969 | Buf1.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 970 | Buf2.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 971 | memset(Buf2.au32Page, 0x77, sizeof(Buf2.au32Page));
|
|---|
| 972 | Buf2.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 973 | Buf3.u64Magic1 = UINT64_C(0xffffffffffffffff);
|
|---|
| 974 | memset(Buf3.au32Page, 0x99, sizeof(Buf3.au32Page));
|
|---|
| 975 | Buf3.u64Magic2 = UINT64_C(0xffffffffffffffff);
|
|---|
| 976 | ASMMemFill32(Buf1.au32Page, sizeof(Buf1.au32Page), 0xdeadbeef);
|
|---|
| 977 | ASMMemFill32(Buf2.au32Page, sizeof(Buf2.au32Page), 0xcafeff01);
|
|---|
| 978 | ASMMemFill32(Buf3.au32Page, sizeof(Buf3.au32Page), 0xf00dd00f);
|
|---|
| 979 | if ( Buf1.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 980 | || Buf1.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 981 | || Buf2.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 982 | || Buf2.u64Magic2 != UINT64_C(0xffffffffffffffff)
|
|---|
| 983 | || Buf3.u64Magic1 != UINT64_C(0xffffffffffffffff)
|
|---|
| 984 | || Buf3.u64Magic2 != UINT64_C(0xffffffffffffffff))
|
|---|
| 985 | {
|
|---|
| 986 | RTPrintf("tstInlineAsm: ASMMemFill32 violated one/both magic(s)!\n");
|
|---|
| 987 | RTTestIErrorInc();
|
|---|
| 988 | }
|
|---|
| 989 | for (unsigned i = 0; i < RT_ELEMENTS(Buf1.au32Page); i++)
|
|---|
| 990 | if (Buf1.au32Page[i] != 0xdeadbeef)
|
|---|
| 991 | {
|
|---|
| 992 | RTPrintf("tstInlineAsm: ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf1.au32Page[i], 0xdeadbeef);
|
|---|
| 993 | RTTestIErrorInc();
|
|---|
| 994 | }
|
|---|
| 995 | for (unsigned i = 0; i < RT_ELEMENTS(Buf2.au32Page); i++)
|
|---|
| 996 | if (Buf2.au32Page[i] != 0xcafeff01)
|
|---|
| 997 | {
|
|---|
| 998 | RTPrintf("tstInlineAsm: ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf2.au32Page[i], 0xcafeff01);
|
|---|
| 999 | RTTestIErrorInc();
|
|---|
| 1000 | }
|
|---|
| 1001 | for (unsigned i = 0; i < RT_ELEMENTS(Buf3.au32Page); i++)
|
|---|
| 1002 | if (Buf3.au32Page[i] != 0xf00dd00f)
|
|---|
| 1003 | {
|
|---|
| 1004 | RTPrintf("tstInlineAsm: ASMMemFill32 %#x: %#x exepcted %#x\n", i, Buf3.au32Page[i], 0xf00dd00f);
|
|---|
| 1005 | RTTestIErrorInc();
|
|---|
| 1006 | }
|
|---|
| 1007 | }
|
|---|
| 1008 |
|
|---|
| 1009 |
|
|---|
| 1010 |
|
|---|
| 1011 | void tstASMMath(void)
|
|---|
| 1012 | {
|
|---|
| 1013 | uint64_t u64 = ASMMult2xU32RetU64(UINT32_C(0x80000000), UINT32_C(0x10000000));
|
|---|
| 1014 | CHECKVAL(u64, UINT64_C(0x0800000000000000), "%#018RX64");
|
|---|
| 1015 |
|
|---|
| 1016 | uint32_t u32 = ASMDivU64ByU32RetU32(UINT64_C(0x0800000000000000), UINT32_C(0x10000000));
|
|---|
| 1017 | CHECKVAL(u32, UINT32_C(0x80000000), "%#010RX32");
|
|---|
| 1018 |
|
|---|
| 1019 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000000000001), UINT32_C(0x00000001), UINT32_C(0x00000001));
|
|---|
| 1020 | CHECKVAL(u64, UINT64_C(0x0000000000000001), "%#018RX64");
|
|---|
| 1021 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x0000000100000000), UINT32_C(0x80000000), UINT32_C(0x00000002));
|
|---|
| 1022 | CHECKVAL(u64, UINT64_C(0x4000000000000000), "%#018RX64");
|
|---|
| 1023 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfedcba9876543210), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
|
|---|
| 1024 | CHECKVAL(u64, UINT64_C(0xfedcba9876543210), "%#018RX64");
|
|---|
| 1025 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xffffffff), UINT32_C(0xffffffff));
|
|---|
| 1026 | CHECKVAL(u64, UINT64_C(0xffffffffffffffff), "%#018RX64");
|
|---|
| 1027 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xffffffffffffffff), UINT32_C(0xfffffff0), UINT32_C(0xffffffff));
|
|---|
| 1028 | CHECKVAL(u64, UINT64_C(0xfffffff0fffffff0), "%#018RX64");
|
|---|
| 1029 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0x58734981), UINT32_C(0xf8694045));
|
|---|
| 1030 | CHECKVAL(u64, UINT64_C(0x128b9c3d43184763), "%#018RX64");
|
|---|
| 1031 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0x3415934810359583), UINT32_C(0xf8694045), UINT32_C(0x58734981));
|
|---|
| 1032 | CHECKVAL(u64, UINT64_C(0x924719355cd35a27), "%#018RX64");
|
|---|
| 1033 |
|
|---|
| 1034 | #if 0 /* bird: question is whether this should trap or not:
|
|---|
| 1035 | *
|
|---|
| 1036 | * frank: Of course it must trap:
|
|---|
| 1037 | *
|
|---|
| 1038 | * 0xfffffff8 * 0x77d7daf8 = 0x77d7daf441412840
|
|---|
| 1039 | *
|
|---|
| 1040 | * During the following division, the quotient must fit into a 32-bit register.
|
|---|
| 1041 | * Therefore the smallest valid divisor is
|
|---|
| 1042 | *
|
|---|
| 1043 | * (0x77d7daf441412840 >> 32) + 1 = 0x77d7daf5
|
|---|
| 1044 | *
|
|---|
| 1045 | * which is definitely greater than 0x3b9aca00.
|
|---|
| 1046 | *
|
|---|
| 1047 | * bird: No, the C version does *not* crash. So, the question is whether there's any
|
|---|
| 1048 | * code depending on it not crashing.
|
|---|
| 1049 | *
|
|---|
| 1050 | * Of course the assembly versions of the code crash right now for the reasons you've
|
|---|
| 1051 | * given, but the 32-bit MSC version does not crash.
|
|---|
| 1052 | *
|
|---|
| 1053 | * frank: The C version does not crash but delivers incorrect results for this case.
|
|---|
| 1054 | * The reason is
|
|---|
| 1055 | *
|
|---|
| 1056 | * u.s.Hi = (unsigned long)(u64Hi / u32C);
|
|---|
| 1057 | *
|
|---|
| 1058 | * Here the division is actually 64-bit by 64-bit but the 64-bit result is truncated
|
|---|
| 1059 | * to 32 bit. If using this (optimized and fast) function we should just be sure that
|
|---|
| 1060 | * the operands are in a valid range.
|
|---|
| 1061 | */
|
|---|
| 1062 | u64 = ASMMultU64ByU32DivByU32(UINT64_C(0xfffffff8c65d6731), UINT32_C(0x77d7daf8), UINT32_C(0x3b9aca00));
|
|---|
| 1063 | CHECKVAL(u64, UINT64_C(0x02b8f9a2aa74e3dc), "%#018RX64");
|
|---|
| 1064 | #endif
|
|---|
| 1065 |
|
|---|
| 1066 | u32 = ASMModU64ByU32RetU32(UINT64_C(0x0ffffff8c65d6731), UINT32_C(0x77d7daf8));
|
|---|
| 1067 | CHECKVAL(u32, UINT32_C(0x3B642451), "%#010RX32");
|
|---|
| 1068 |
|
|---|
| 1069 | int32_t i32;
|
|---|
| 1070 | i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(-2));
|
|---|
| 1071 | CHECKVAL(i32, INT32_C(-1), "%010RI32");
|
|---|
| 1072 | i32 = ASMModS64ByS32RetS32(INT64_C(-11), INT32_C(2));
|
|---|
| 1073 | CHECKVAL(i32, INT32_C(-1), "%010RI32");
|
|---|
| 1074 | i32 = ASMModS64ByS32RetS32(INT64_C(11), INT32_C(-2));
|
|---|
| 1075 | CHECKVAL(i32, INT32_C(1), "%010RI32");
|
|---|
| 1076 |
|
|---|
| 1077 | i32 = ASMModS64ByS32RetS32(INT64_C(92233720368547758), INT32_C(2147483647));
|
|---|
| 1078 | CHECKVAL(i32, INT32_C(2104533974), "%010RI32");
|
|---|
| 1079 | i32 = ASMModS64ByS32RetS32(INT64_C(-92233720368547758), INT32_C(2147483647));
|
|---|
| 1080 | CHECKVAL(i32, INT32_C(-2104533974), "%010RI32");
|
|---|
| 1081 | }
|
|---|
| 1082 |
|
|---|
| 1083 |
|
|---|
| 1084 | void tstASMByteSwap(void)
|
|---|
| 1085 | {
|
|---|
| 1086 | RTPrintf("tstInlineASM: TESTING - ASMByteSwap*\n");
|
|---|
| 1087 |
|
|---|
| 1088 | uint64_t u64In = UINT64_C(0x0011223344556677);
|
|---|
| 1089 | uint64_t u64Out = ASMByteSwapU64(u64In);
|
|---|
| 1090 | CHECKVAL(u64In, UINT64_C(0x0011223344556677), "%#018RX64");
|
|---|
| 1091 | CHECKVAL(u64Out, UINT64_C(0x7766554433221100), "%#018RX64");
|
|---|
| 1092 | u64Out = ASMByteSwapU64(u64Out);
|
|---|
| 1093 | CHECKVAL(u64Out, u64In, "%#018RX64");
|
|---|
| 1094 | u64In = UINT64_C(0x0123456789abcdef);
|
|---|
| 1095 | u64Out = ASMByteSwapU64(u64In);
|
|---|
| 1096 | CHECKVAL(u64In, UINT64_C(0x0123456789abcdef), "%#018RX64");
|
|---|
| 1097 | CHECKVAL(u64Out, UINT64_C(0xefcdab8967452301), "%#018RX64");
|
|---|
| 1098 | u64Out = ASMByteSwapU64(u64Out);
|
|---|
| 1099 | CHECKVAL(u64Out, u64In, "%#018RX64");
|
|---|
| 1100 | u64In = 0;
|
|---|
| 1101 | u64Out = ASMByteSwapU64(u64In);
|
|---|
| 1102 | CHECKVAL(u64Out, u64In, "%#018RX64");
|
|---|
| 1103 | u64In = ~(uint64_t)0;
|
|---|
| 1104 | u64Out = ASMByteSwapU64(u64In);
|
|---|
| 1105 | CHECKVAL(u64Out, u64In, "%#018RX64");
|
|---|
| 1106 |
|
|---|
| 1107 | uint32_t u32In = UINT32_C(0x00112233);
|
|---|
| 1108 | uint32_t u32Out = ASMByteSwapU32(u32In);
|
|---|
| 1109 | CHECKVAL(u32In, UINT32_C(0x00112233), "%#010RX32");
|
|---|
| 1110 | CHECKVAL(u32Out, UINT32_C(0x33221100), "%#010RX32");
|
|---|
| 1111 | u32Out = ASMByteSwapU32(u32Out);
|
|---|
| 1112 | CHECKVAL(u32Out, u32In, "%#010RX32");
|
|---|
| 1113 | u32In = UINT32_C(0x12345678);
|
|---|
| 1114 | u32Out = ASMByteSwapU32(u32In);
|
|---|
| 1115 | CHECKVAL(u32In, UINT32_C(0x12345678), "%#010RX32");
|
|---|
| 1116 | CHECKVAL(u32Out, UINT32_C(0x78563412), "%#010RX32");
|
|---|
| 1117 | u32Out = ASMByteSwapU32(u32Out);
|
|---|
| 1118 | CHECKVAL(u32Out, u32In, "%#010RX32");
|
|---|
| 1119 | u32In = 0;
|
|---|
| 1120 | u32Out = ASMByteSwapU32(u32In);
|
|---|
| 1121 | CHECKVAL(u32Out, u32In, "%#010RX32");
|
|---|
| 1122 | u32In = ~(uint32_t)0;
|
|---|
| 1123 | u32Out = ASMByteSwapU32(u32In);
|
|---|
| 1124 | CHECKVAL(u32Out, u32In, "%#010RX32");
|
|---|
| 1125 |
|
|---|
| 1126 | uint16_t u16In = UINT16_C(0x0011);
|
|---|
| 1127 | uint16_t u16Out = ASMByteSwapU16(u16In);
|
|---|
| 1128 | CHECKVAL(u16In, UINT16_C(0x0011), "%#06RX16");
|
|---|
| 1129 | CHECKVAL(u16Out, UINT16_C(0x1100), "%#06RX16");
|
|---|
| 1130 | u16Out = ASMByteSwapU16(u16Out);
|
|---|
| 1131 | CHECKVAL(u16Out, u16In, "%#06RX16");
|
|---|
| 1132 | u16In = UINT16_C(0x1234);
|
|---|
| 1133 | u16Out = ASMByteSwapU16(u16In);
|
|---|
| 1134 | CHECKVAL(u16In, UINT16_C(0x1234), "%#06RX16");
|
|---|
| 1135 | CHECKVAL(u16Out, UINT16_C(0x3412), "%#06RX16");
|
|---|
| 1136 | u16Out = ASMByteSwapU16(u16Out);
|
|---|
| 1137 | CHECKVAL(u16Out, u16In, "%#06RX16");
|
|---|
| 1138 | u16In = 0;
|
|---|
| 1139 | u16Out = ASMByteSwapU16(u16In);
|
|---|
| 1140 | CHECKVAL(u16Out, u16In, "%#06RX16");
|
|---|
| 1141 | u16In = ~(uint16_t)0;
|
|---|
| 1142 | u16Out = ASMByteSwapU16(u16In);
|
|---|
| 1143 | CHECKVAL(u16Out, u16In, "%#06RX16");
|
|---|
| 1144 | }
|
|---|
| 1145 |
|
|---|
| 1146 |
|
|---|
| 1147 | void tstASMBench(void)
|
|---|
| 1148 | {
|
|---|
| 1149 | /*
|
|---|
| 1150 | * Make this static. We don't want to have this located on the stack.
|
|---|
| 1151 | */
|
|---|
| 1152 | static uint8_t volatile s_u8;
|
|---|
| 1153 | static int8_t volatile s_i8;
|
|---|
| 1154 | static uint16_t volatile s_u16;
|
|---|
| 1155 | static int16_t volatile s_i16;
|
|---|
| 1156 | static uint32_t volatile s_u32;
|
|---|
| 1157 | static int32_t volatile s_i32;
|
|---|
| 1158 | static uint64_t volatile s_u64;
|
|---|
| 1159 | static int64_t volatile s_i64;
|
|---|
| 1160 | register unsigned i;
|
|---|
| 1161 | const unsigned cRounds = 1000000;
|
|---|
| 1162 | register uint64_t u64Elapsed;
|
|---|
| 1163 |
|
|---|
| 1164 | RTPrintf("tstInlineASM: Benchmarking:\n");
|
|---|
| 1165 |
|
|---|
| 1166 | #define BENCH(op, str) \
|
|---|
| 1167 | RTThreadYield(); \
|
|---|
| 1168 | u64Elapsed = ASMReadTSC(); \
|
|---|
| 1169 | for (i = cRounds; i > 0; i--) \
|
|---|
| 1170 | op; \
|
|---|
| 1171 | u64Elapsed = ASMReadTSC() - u64Elapsed; \
|
|---|
| 1172 | RTPrintf(" %-30s %3llu cycles\n", str, u64Elapsed / cRounds);
|
|---|
| 1173 |
|
|---|
| 1174 | BENCH(s_u32 = 0, "s_u32 = 0:");
|
|---|
| 1175 | BENCH(ASMAtomicUoWriteU8(&s_u8, 0), "ASMAtomicUoWriteU8:");
|
|---|
| 1176 | BENCH(ASMAtomicUoWriteS8(&s_i8, 0), "ASMAtomicUoWriteS8:");
|
|---|
| 1177 | BENCH(ASMAtomicUoWriteU16(&s_u16, 0), "ASMAtomicUoWriteU16:");
|
|---|
| 1178 | BENCH(ASMAtomicUoWriteS16(&s_i16, 0), "ASMAtomicUoWriteS16:");
|
|---|
| 1179 | BENCH(ASMAtomicUoWriteU32(&s_u32, 0), "ASMAtomicUoWriteU32:");
|
|---|
| 1180 | BENCH(ASMAtomicUoWriteS32(&s_i32, 0), "ASMAtomicUoWriteS32:");
|
|---|
| 1181 | BENCH(ASMAtomicUoWriteU64(&s_u64, 0), "ASMAtomicUoWriteU64:");
|
|---|
| 1182 | BENCH(ASMAtomicUoWriteS64(&s_i64, 0), "ASMAtomicUoWriteS64:");
|
|---|
| 1183 | BENCH(ASMAtomicWriteU8(&s_u8, 0), "ASMAtomicWriteU8:");
|
|---|
| 1184 | BENCH(ASMAtomicWriteS8(&s_i8, 0), "ASMAtomicWriteS8:");
|
|---|
| 1185 | BENCH(ASMAtomicWriteU16(&s_u16, 0), "ASMAtomicWriteU16:");
|
|---|
| 1186 | BENCH(ASMAtomicWriteS16(&s_i16, 0), "ASMAtomicWriteS16:");
|
|---|
| 1187 | BENCH(ASMAtomicWriteU32(&s_u32, 0), "ASMAtomicWriteU32:");
|
|---|
| 1188 | BENCH(ASMAtomicWriteS32(&s_i32, 0), "ASMAtomicWriteS32:");
|
|---|
| 1189 | BENCH(ASMAtomicWriteU64(&s_u64, 0), "ASMAtomicWriteU64:");
|
|---|
| 1190 | BENCH(ASMAtomicWriteS64(&s_i64, 0), "ASMAtomicWriteS64:");
|
|---|
| 1191 | BENCH(ASMAtomicXchgU8(&s_u8, 0), "ASMAtomicXchgU8:");
|
|---|
| 1192 | BENCH(ASMAtomicXchgS8(&s_i8, 0), "ASMAtomicXchgS8:");
|
|---|
| 1193 | BENCH(ASMAtomicXchgU16(&s_u16, 0), "ASMAtomicXchgU16:");
|
|---|
| 1194 | BENCH(ASMAtomicXchgS16(&s_i16, 0), "ASMAtomicXchgS16:");
|
|---|
| 1195 | BENCH(ASMAtomicXchgU32(&s_u32, 0), "ASMAtomicXchgU32:");
|
|---|
| 1196 | BENCH(ASMAtomicXchgS32(&s_i32, 0), "ASMAtomicXchgS32:");
|
|---|
| 1197 | BENCH(ASMAtomicXchgU64(&s_u64, 0), "ASMAtomicXchgU64:");
|
|---|
| 1198 | BENCH(ASMAtomicXchgS64(&s_i64, 0), "ASMAtomicXchgS64:");
|
|---|
| 1199 | BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 0), "ASMAtomicCmpXchgU32:");
|
|---|
| 1200 | BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 0), "ASMAtomicCmpXchgS32:");
|
|---|
| 1201 | BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 0), "ASMAtomicCmpXchgU64:");
|
|---|
| 1202 | BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 0), "ASMAtomicCmpXchgS64:");
|
|---|
| 1203 | BENCH(ASMAtomicCmpXchgU32(&s_u32, 0, 1), "ASMAtomicCmpXchgU32/neg:");
|
|---|
| 1204 | BENCH(ASMAtomicCmpXchgS32(&s_i32, 0, 1), "ASMAtomicCmpXchgS32/neg:");
|
|---|
| 1205 | BENCH(ASMAtomicCmpXchgU64(&s_u64, 0, 1), "ASMAtomicCmpXchgU64/neg:");
|
|---|
| 1206 | BENCH(ASMAtomicCmpXchgS64(&s_i64, 0, 1), "ASMAtomicCmpXchgS64/neg:");
|
|---|
| 1207 | BENCH(ASMAtomicIncU32(&s_u32), "ASMAtomicIncU32:");
|
|---|
| 1208 | BENCH(ASMAtomicIncS32(&s_i32), "ASMAtomicIncS32:");
|
|---|
| 1209 | BENCH(ASMAtomicDecU32(&s_u32), "ASMAtomicDecU32:");
|
|---|
| 1210 | BENCH(ASMAtomicDecS32(&s_i32), "ASMAtomicDecS32:");
|
|---|
| 1211 | BENCH(ASMAtomicAddU32(&s_u32, 5), "ASMAtomicAddU32:");
|
|---|
| 1212 | BENCH(ASMAtomicAddS32(&s_i32, 5), "ASMAtomicAddS32:");
|
|---|
| 1213 |
|
|---|
| 1214 | RTPrintf("Done.\n");
|
|---|
| 1215 |
|
|---|
| 1216 | #undef BENCH
|
|---|
| 1217 | }
|
|---|
| 1218 |
|
|---|
| 1219 |
|
|---|
| 1220 | int main(int argc, char *argv[])
|
|---|
| 1221 | {
|
|---|
| 1222 | RTTEST hTest;
|
|---|
| 1223 | int rc = RTTestInitAndCreate("tstInlineAsm", &hTest);
|
|---|
| 1224 | if (rc)
|
|---|
| 1225 | return rc;
|
|---|
| 1226 | RTTestBanner(hTest);
|
|---|
| 1227 |
|
|---|
| 1228 | /*
|
|---|
| 1229 | * Execute the tests.
|
|---|
| 1230 | */
|
|---|
| 1231 | #if !defined(PIC) || !defined(RT_ARCH_X86)
|
|---|
| 1232 | tstASMCpuId();
|
|---|
| 1233 | #endif
|
|---|
| 1234 | tstASMAtomicXchgU8();
|
|---|
| 1235 | tstASMAtomicXchgU16();
|
|---|
| 1236 | tstASMAtomicXchgU32();
|
|---|
| 1237 | tstASMAtomicXchgU64();
|
|---|
| 1238 | tstASMAtomicXchgPtr();
|
|---|
| 1239 | tstASMAtomicCmpXchgU8();
|
|---|
| 1240 | tstASMAtomicCmpXchgU32();
|
|---|
| 1241 | tstASMAtomicCmpXchgU64();
|
|---|
| 1242 | tstASMAtomicCmpXchgExU32();
|
|---|
| 1243 | tstASMAtomicCmpXchgExU64();
|
|---|
| 1244 | tstASMAtomicReadU64();
|
|---|
| 1245 | tstASMAtomicAddS32();
|
|---|
| 1246 | tstASMAtomicDecIncS32();
|
|---|
| 1247 | tstASMAtomicAndOrU32();
|
|---|
| 1248 | tstASMMemZeroPage();
|
|---|
| 1249 | tstASMMemIsZeroPage(hTest);
|
|---|
| 1250 | tstASMMemZero32();
|
|---|
| 1251 | tstASMMemFill32();
|
|---|
| 1252 | tstASMMath();
|
|---|
| 1253 | tstASMByteSwap();
|
|---|
| 1254 | tstASMBench();
|
|---|
| 1255 |
|
|---|
| 1256 | /*
|
|---|
| 1257 | * Show the result.
|
|---|
| 1258 | */
|
|---|
| 1259 | return RTTestSummaryAndDestroy(hTest);
|
|---|
| 1260 | }
|
|---|
| 1261 |
|
|---|
