1 //===-- Host.cpp - Implement OS Host Concept --------------------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file implements the operating system Host concept. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Support/Host.h" 15 #include "llvm/Support/TargetParser.h" 16 #include "llvm/ADT/SmallSet.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/StringSwitch.h" 20 #include "llvm/ADT/Triple.h" 21 #include "llvm/Config/llvm-config.h" 22 #include "llvm/Support/Debug.h" 23 #include "llvm/Support/FileSystem.h" 24 #include "llvm/Support/MemoryBuffer.h" 25 #include "llvm/Support/raw_ostream.h" 26 #include <assert.h> 27 #include <string.h> 28 29 // Include the platform-specific parts of this class. 30 #ifdef LLVM_ON_UNIX 31 #include "Unix/Host.inc" 32 #endif 33 #ifdef _WIN32 34 #include "Windows/Host.inc" 35 #endif 36 #ifdef _MSC_VER 37 #include <intrin.h> 38 #endif 39 #if defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__)) 40 #include <mach/host_info.h> 41 #include <mach/mach.h> 42 #include <mach/mach_host.h> 43 #include <mach/machine.h> 44 #endif 45 46 #define DEBUG_TYPE "host-detection" 47 48 //===----------------------------------------------------------------------===// 49 // 50 // Implementations of the CPU detection routines 51 // 52 //===----------------------------------------------------------------------===// 53 54 using namespace llvm; 55 56 static std::unique_ptr<llvm::MemoryBuffer> 57 LLVM_ATTRIBUTE_UNUSED getProcCpuinfoContent() { 58 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> Text = 59 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo"); 60 if (std::error_code EC = Text.getError()) { 61 llvm::errs() << "Can't read " 62 << "/proc/cpuinfo: " << EC.message() << "\n"; 63 return nullptr; 64 } 65 return std::move(*Text); 66 } 67 68 StringRef sys::detail::getHostCPUNameForPowerPC(StringRef ProcCpuinfoContent) { 69 // Access to the Processor Version Register (PVR) on PowerPC is privileged, 70 // and so we must use an operating-system interface to determine the current 71 // processor type. On Linux, this is exposed through the /proc/cpuinfo file. 72 const char *generic = "generic"; 73 74 // The cpu line is second (after the 'processor: 0' line), so if this 75 // buffer is too small then something has changed (or is wrong). 76 StringRef::const_iterator CPUInfoStart = ProcCpuinfoContent.begin(); 77 StringRef::const_iterator CPUInfoEnd = ProcCpuinfoContent.end(); 78 79 StringRef::const_iterator CIP = CPUInfoStart; 80 81 StringRef::const_iterator CPUStart = 0; 82 size_t CPULen = 0; 83 84 // We need to find the first line which starts with cpu, spaces, and a colon. 85 // After the colon, there may be some additional spaces and then the cpu type. 86 while (CIP < CPUInfoEnd && CPUStart == 0) { 87 if (CIP < CPUInfoEnd && *CIP == '\n') 88 ++CIP; 89 90 if (CIP < CPUInfoEnd && *CIP == 'c') { 91 ++CIP; 92 if (CIP < CPUInfoEnd && *CIP == 'p') { 93 ++CIP; 94 if (CIP < CPUInfoEnd && *CIP == 'u') { 95 ++CIP; 96 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t')) 97 ++CIP; 98 99 if (CIP < CPUInfoEnd && *CIP == ':') { 100 ++CIP; 101 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t')) 102 ++CIP; 103 104 if (CIP < CPUInfoEnd) { 105 CPUStart = CIP; 106 while (CIP < CPUInfoEnd && (*CIP != ' ' && *CIP != '\t' && 107 *CIP != ',' && *CIP != '\n')) 108 ++CIP; 109 CPULen = CIP - CPUStart; 110 } 111 } 112 } 113 } 114 } 115 116 if (CPUStart == 0) 117 while (CIP < CPUInfoEnd && *CIP != '\n') 118 ++CIP; 119 } 120 121 if (CPUStart == 0) 122 return generic; 123 124 return StringSwitch<const char *>(StringRef(CPUStart, CPULen)) 125 .Case("604e", "604e") 126 .Case("604", "604") 127 .Case("7400", "7400") 128 .Case("7410", "7400") 129 .Case("7447", "7400") 130 .Case("7455", "7450") 131 .Case("G4", "g4") 132 .Case("POWER4", "970") 133 .Case("PPC970FX", "970") 134 .Case("PPC970MP", "970") 135 .Case("G5", "g5") 136 .Case("POWER5", "g5") 137 .Case("A2", "a2") 138 .Case("POWER6", "pwr6") 139 .Case("POWER7", "pwr7") 140 .Case("POWER8", "pwr8") 141 .Case("POWER8E", "pwr8") 142 .Case("POWER8NVL", "pwr8") 143 .Case("POWER9", "pwr9") 144 .Default(generic); 145 } 146 147 StringRef sys::detail::getHostCPUNameForARM(StringRef ProcCpuinfoContent) { 148 // The cpuid register on arm is not accessible from user space. On Linux, 149 // it is exposed through the /proc/cpuinfo file. 150 151 // Read 32 lines from /proc/cpuinfo, which should contain the CPU part line 152 // in all cases. 153 SmallVector<StringRef, 32> Lines; 154 ProcCpuinfoContent.split(Lines, "\n"); 155 156 // Look for the CPU implementer line. 157 StringRef Implementer; 158 StringRef Hardware; 159 for (unsigned I = 0, E = Lines.size(); I != E; ++I) { 160 if (Lines[I].startswith("CPU implementer")) 161 Implementer = Lines[I].substr(15).ltrim("\t :"); 162 if (Lines[I].startswith("Hardware")) 163 Hardware = Lines[I].substr(8).ltrim("\t :"); 164 } 165 166 if (Implementer == "0x41") { // ARM Ltd. 167 // MSM8992/8994 may give cpu part for the core that the kernel is running on, 168 // which is undeterministic and wrong. Always return cortex-a53 for these SoC. 169 if (Hardware.endswith("MSM8994") || Hardware.endswith("MSM8996")) 170 return "cortex-a53"; 171 172 173 // Look for the CPU part line. 174 for (unsigned I = 0, E = Lines.size(); I != E; ++I) 175 if (Lines[I].startswith("CPU part")) 176 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The 177 // values correspond to the "Part number" in the CP15/c0 register. The 178 // contents are specified in the various processor manuals. 179 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :")) 180 .Case("0x926", "arm926ej-s") 181 .Case("0xb02", "mpcore") 182 .Case("0xb36", "arm1136j-s") 183 .Case("0xb56", "arm1156t2-s") 184 .Case("0xb76", "arm1176jz-s") 185 .Case("0xc08", "cortex-a8") 186 .Case("0xc09", "cortex-a9") 187 .Case("0xc0f", "cortex-a15") 188 .Case("0xc20", "cortex-m0") 189 .Case("0xc23", "cortex-m3") 190 .Case("0xc24", "cortex-m4") 191 .Case("0xd04", "cortex-a35") 192 .Case("0xd03", "cortex-a53") 193 .Case("0xd07", "cortex-a57") 194 .Case("0xd08", "cortex-a72") 195 .Case("0xd09", "cortex-a73") 196 .Default("generic"); 197 } 198 199 if (Implementer == "0x42" || Implementer == "0x43") { // Broadcom | Cavium. 200 for (unsigned I = 0, E = Lines.size(); I != E; ++I) { 201 if (Lines[I].startswith("CPU part")) { 202 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :")) 203 .Case("0x516", "thunderx2t99") 204 .Case("0x0516", "thunderx2t99") 205 .Case("0xaf", "thunderx2t99") 206 .Case("0x0af", "thunderx2t99") 207 .Case("0xa1", "thunderxt88") 208 .Case("0x0a1", "thunderxt88") 209 .Default("generic"); 210 } 211 } 212 } 213 214 if (Implementer == "0x48") // HiSilicon Technologies, Inc. 215 // Look for the CPU part line. 216 for (unsigned I = 0, E = Lines.size(); I != E; ++I) 217 if (Lines[I].startswith("CPU part")) 218 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The 219 // values correspond to the "Part number" in the CP15/c0 register. The 220 // contents are specified in the various processor manuals. 221 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :")) 222 .Case("0xd01", "tsv110") 223 .Default("generic"); 224 225 if (Implementer == "0x51") // Qualcomm Technologies, Inc. 226 // Look for the CPU part line. 227 for (unsigned I = 0, E = Lines.size(); I != E; ++I) 228 if (Lines[I].startswith("CPU part")) 229 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The 230 // values correspond to the "Part number" in the CP15/c0 register. The 231 // contents are specified in the various processor manuals. 232 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :")) 233 .Case("0x06f", "krait") // APQ8064 234 .Case("0x201", "kryo") 235 .Case("0x205", "kryo") 236 .Case("0x211", "kryo") 237 .Case("0x800", "cortex-a73") 238 .Case("0x801", "cortex-a73") 239 .Case("0xc00", "falkor") 240 .Case("0xc01", "saphira") 241 .Default("generic"); 242 243 if (Implementer == "0x53") { // Samsung Electronics Co., Ltd. 244 // The Exynos chips have a convoluted ID scheme that doesn't seem to follow 245 // any predictive pattern across variants and parts. 246 unsigned Variant = 0, Part = 0; 247 248 // Look for the CPU variant line, whose value is a 1 digit hexadecimal 249 // number, corresponding to the Variant bits in the CP15/C0 register. 250 for (auto I : Lines) 251 if (I.consume_front("CPU variant")) 252 I.ltrim("\t :").getAsInteger(0, Variant); 253 254 // Look for the CPU part line, whose value is a 3 digit hexadecimal 255 // number, corresponding to the PartNum bits in the CP15/C0 register. 256 for (auto I : Lines) 257 if (I.consume_front("CPU part")) 258 I.ltrim("\t :").getAsInteger(0, Part); 259 260 unsigned Exynos = (Variant << 12) | Part; 261 switch (Exynos) { 262 default: 263 // Default by falling through to Exynos M1. 264 LLVM_FALLTHROUGH; 265 266 case 0x1001: 267 return "exynos-m1"; 268 269 case 0x4001: 270 return "exynos-m2"; 271 } 272 } 273 274 return "generic"; 275 } 276 277 StringRef sys::detail::getHostCPUNameForS390x(StringRef ProcCpuinfoContent) { 278 // STIDP is a privileged operation, so use /proc/cpuinfo instead. 279 280 // The "processor 0:" line comes after a fair amount of other information, 281 // including a cache breakdown, but this should be plenty. 282 SmallVector<StringRef, 32> Lines; 283 ProcCpuinfoContent.split(Lines, "\n"); 284 285 // Look for the CPU features. 286 SmallVector<StringRef, 32> CPUFeatures; 287 for (unsigned I = 0, E = Lines.size(); I != E; ++I) 288 if (Lines[I].startswith("features")) { 289 size_t Pos = Lines[I].find(":"); 290 if (Pos != StringRef::npos) { 291 Lines[I].drop_front(Pos + 1).split(CPUFeatures, ' '); 292 break; 293 } 294 } 295 296 // We need to check for the presence of vector support independently of 297 // the machine type, since we may only use the vector register set when 298 // supported by the kernel (and hypervisor). 299 bool HaveVectorSupport = false; 300 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) { 301 if (CPUFeatures[I] == "vx") 302 HaveVectorSupport = true; 303 } 304 305 // Now check the processor machine type. 306 for (unsigned I = 0, E = Lines.size(); I != E; ++I) { 307 if (Lines[I].startswith("processor ")) { 308 size_t Pos = Lines[I].find("machine = "); 309 if (Pos != StringRef::npos) { 310 Pos += sizeof("machine = ") - 1; 311 unsigned int Id; 312 if (!Lines[I].drop_front(Pos).getAsInteger(10, Id)) { 313 if (Id >= 3906 && HaveVectorSupport) 314 return "z14"; 315 if (Id >= 2964 && HaveVectorSupport) 316 return "z13"; 317 if (Id >= 2827) 318 return "zEC12"; 319 if (Id >= 2817) 320 return "z196"; 321 } 322 } 323 break; 324 } 325 } 326 327 return "generic"; 328 } 329 330 StringRef sys::detail::getHostCPUNameForBPF() { 331 #if !defined(__linux__) || !defined(__x86_64__) 332 return "generic"; 333 #else 334 uint8_t insns[40] __attribute__ ((aligned (8))) = 335 /* BPF_MOV64_IMM(BPF_REG_0, 0) */ 336 { 0xb7, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 337 /* BPF_MOV64_IMM(BPF_REG_2, 1) */ 338 0xb7, 0x2, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0, 339 /* BPF_JMP_REG(BPF_JLT, BPF_REG_0, BPF_REG_2, 1) */ 340 0xad, 0x20, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 341 /* BPF_MOV64_IMM(BPF_REG_0, 1) */ 342 0xb7, 0x0, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0, 343 /* BPF_EXIT_INSN() */ 344 0x95, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 }; 345 346 struct bpf_prog_load_attr { 347 uint32_t prog_type; 348 uint32_t insn_cnt; 349 uint64_t insns; 350 uint64_t license; 351 uint32_t log_level; 352 uint32_t log_size; 353 uint64_t log_buf; 354 uint32_t kern_version; 355 uint32_t prog_flags; 356 } attr = {}; 357 attr.prog_type = 1; /* BPF_PROG_TYPE_SOCKET_FILTER */ 358 attr.insn_cnt = 5; 359 attr.insns = (uint64_t)insns; 360 attr.license = (uint64_t)"DUMMY"; 361 362 int fd = syscall(321 /* __NR_bpf */, 5 /* BPF_PROG_LOAD */, &attr, sizeof(attr)); 363 if (fd >= 0) { 364 close(fd); 365 return "v2"; 366 } 367 return "v1"; 368 #endif 369 } 370 371 #if defined(__i386__) || defined(_M_IX86) || \ 372 defined(__x86_64__) || defined(_M_X64) 373 374 enum VendorSignatures { 375 SIG_INTEL = 0x756e6547 /* Genu */, 376 SIG_AMD = 0x68747541 /* Auth */ 377 }; 378 379 // The check below for i386 was copied from clang's cpuid.h (__get_cpuid_max). 380 // Check motivated by bug reports for OpenSSL crashing on CPUs without CPUID 381 // support. Consequently, for i386, the presence of CPUID is checked first 382 // via the corresponding eflags bit. 383 // Removal of cpuid.h header motivated by PR30384 384 // Header cpuid.h and method __get_cpuid_max are not used in llvm, clang, openmp 385 // or test-suite, but are used in external projects e.g. libstdcxx 386 static bool isCpuIdSupported() { 387 #if defined(__GNUC__) || defined(__clang__) 388 #if defined(__i386__) 389 int __cpuid_supported; 390 __asm__(" pushfl\n" 391 " popl %%eax\n" 392 " movl %%eax,%%ecx\n" 393 " xorl $0x00200000,%%eax\n" 394 " pushl %%eax\n" 395 " popfl\n" 396 " pushfl\n" 397 " popl %%eax\n" 398 " movl $0,%0\n" 399 " cmpl %%eax,%%ecx\n" 400 " je 1f\n" 401 " movl $1,%0\n" 402 "1:" 403 : "=r"(__cpuid_supported) 404 : 405 : "eax", "ecx"); 406 if (!__cpuid_supported) 407 return false; 408 #endif 409 return true; 410 #endif 411 return true; 412 } 413 414 /// getX86CpuIDAndInfo - Execute the specified cpuid and return the 4 values in 415 /// the specified arguments. If we can't run cpuid on the host, return true. 416 static bool getX86CpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX, 417 unsigned *rECX, unsigned *rEDX) { 418 #if defined(__GNUC__) || defined(__clang__) 419 #if defined(__x86_64__) 420 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually. 421 // FIXME: should we save this for Clang? 422 __asm__("movq\t%%rbx, %%rsi\n\t" 423 "cpuid\n\t" 424 "xchgq\t%%rbx, %%rsi\n\t" 425 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX) 426 : "a"(value)); 427 return false; 428 #elif defined(__i386__) 429 __asm__("movl\t%%ebx, %%esi\n\t" 430 "cpuid\n\t" 431 "xchgl\t%%ebx, %%esi\n\t" 432 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX) 433 : "a"(value)); 434 return false; 435 #else 436 return true; 437 #endif 438 #elif defined(_MSC_VER) 439 // The MSVC intrinsic is portable across x86 and x64. 440 int registers[4]; 441 __cpuid(registers, value); 442 *rEAX = registers[0]; 443 *rEBX = registers[1]; 444 *rECX = registers[2]; 445 *rEDX = registers[3]; 446 return false; 447 #else 448 return true; 449 #endif 450 } 451 452 /// getX86CpuIDAndInfoEx - Execute the specified cpuid with subleaf and return 453 /// the 4 values in the specified arguments. If we can't run cpuid on the host, 454 /// return true. 455 static bool getX86CpuIDAndInfoEx(unsigned value, unsigned subleaf, 456 unsigned *rEAX, unsigned *rEBX, unsigned *rECX, 457 unsigned *rEDX) { 458 #if defined(__GNUC__) || defined(__clang__) 459 #if defined(__x86_64__) 460 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually. 461 // FIXME: should we save this for Clang? 462 __asm__("movq\t%%rbx, %%rsi\n\t" 463 "cpuid\n\t" 464 "xchgq\t%%rbx, %%rsi\n\t" 465 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX) 466 : "a"(value), "c"(subleaf)); 467 return false; 468 #elif defined(__i386__) 469 __asm__("movl\t%%ebx, %%esi\n\t" 470 "cpuid\n\t" 471 "xchgl\t%%ebx, %%esi\n\t" 472 : "=a"(*rEAX), "=S"(*rEBX), "=c"(*rECX), "=d"(*rEDX) 473 : "a"(value), "c"(subleaf)); 474 return false; 475 #else 476 return true; 477 #endif 478 #elif defined(_MSC_VER) 479 int registers[4]; 480 __cpuidex(registers, value, subleaf); 481 *rEAX = registers[0]; 482 *rEBX = registers[1]; 483 *rECX = registers[2]; 484 *rEDX = registers[3]; 485 return false; 486 #else 487 return true; 488 #endif 489 } 490 491 // Read control register 0 (XCR0). Used to detect features such as AVX. 492 static bool getX86XCR0(unsigned *rEAX, unsigned *rEDX) { 493 #if defined(__GNUC__) || defined(__clang__) 494 // Check xgetbv; this uses a .byte sequence instead of the instruction 495 // directly because older assemblers do not include support for xgetbv and 496 // there is no easy way to conditionally compile based on the assembler used. 497 __asm__(".byte 0x0f, 0x01, 0xd0" : "=a"(*rEAX), "=d"(*rEDX) : "c"(0)); 498 return false; 499 #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK) 500 unsigned long long Result = _xgetbv(_XCR_XFEATURE_ENABLED_MASK); 501 *rEAX = Result; 502 *rEDX = Result >> 32; 503 return false; 504 #else 505 return true; 506 #endif 507 } 508 509 static void detectX86FamilyModel(unsigned EAX, unsigned *Family, 510 unsigned *Model) { 511 *Family = (EAX >> 8) & 0xf; // Bits 8 - 11 512 *Model = (EAX >> 4) & 0xf; // Bits 4 - 7 513 if (*Family == 6 || *Family == 0xf) { 514 if (*Family == 0xf) 515 // Examine extended family ID if family ID is F. 516 *Family += (EAX >> 20) & 0xff; // Bits 20 - 27 517 // Examine extended model ID if family ID is 6 or F. 518 *Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19 519 } 520 } 521 522 static void 523 getIntelProcessorTypeAndSubtype(unsigned Family, unsigned Model, 524 unsigned Brand_id, unsigned Features, 525 unsigned Features2, unsigned Features3, 526 unsigned *Type, unsigned *Subtype) { 527 if (Brand_id != 0) 528 return; 529 switch (Family) { 530 case 3: 531 *Type = X86::INTEL_i386; 532 break; 533 case 4: 534 *Type = X86::INTEL_i486; 535 break; 536 case 5: 537 if (Features & (1 << X86::FEATURE_MMX)) { 538 *Type = X86::INTEL_PENTIUM_MMX; 539 break; 540 } 541 *Type = X86::INTEL_PENTIUM; 542 break; 543 case 6: 544 switch (Model) { 545 case 0x01: // Pentium Pro processor 546 *Type = X86::INTEL_PENTIUM_PRO; 547 break; 548 case 0x03: // Intel Pentium II OverDrive processor, Pentium II processor, 549 // model 03 550 case 0x05: // Pentium II processor, model 05, Pentium II Xeon processor, 551 // model 05, and Intel Celeron processor, model 05 552 case 0x06: // Celeron processor, model 06 553 *Type = X86::INTEL_PENTIUM_II; 554 break; 555 case 0x07: // Pentium III processor, model 07, and Pentium III Xeon 556 // processor, model 07 557 case 0x08: // Pentium III processor, model 08, Pentium III Xeon processor, 558 // model 08, and Celeron processor, model 08 559 case 0x0a: // Pentium III Xeon processor, model 0Ah 560 case 0x0b: // Pentium III processor, model 0Bh 561 *Type = X86::INTEL_PENTIUM_III; 562 break; 563 case 0x09: // Intel Pentium M processor, Intel Celeron M processor model 09. 564 case 0x0d: // Intel Pentium M processor, Intel Celeron M processor, model 565 // 0Dh. All processors are manufactured using the 90 nm process. 566 case 0x15: // Intel EP80579 Integrated Processor and Intel EP80579 567 // Integrated Processor with Intel QuickAssist Technology 568 *Type = X86::INTEL_PENTIUM_M; 569 break; 570 case 0x0e: // Intel Core Duo processor, Intel Core Solo processor, model 571 // 0Eh. All processors are manufactured using the 65 nm process. 572 *Type = X86::INTEL_CORE_DUO; 573 break; // yonah 574 case 0x0f: // Intel Core 2 Duo processor, Intel Core 2 Duo mobile 575 // processor, Intel Core 2 Quad processor, Intel Core 2 Quad 576 // mobile processor, Intel Core 2 Extreme processor, Intel 577 // Pentium Dual-Core processor, Intel Xeon processor, model 578 // 0Fh. All processors are manufactured using the 65 nm process. 579 case 0x16: // Intel Celeron processor model 16h. All processors are 580 // manufactured using the 65 nm process 581 *Type = X86::INTEL_CORE2; // "core2" 582 *Subtype = X86::INTEL_CORE2_65; 583 break; 584 case 0x17: // Intel Core 2 Extreme processor, Intel Xeon processor, model 585 // 17h. All processors are manufactured using the 45 nm process. 586 // 587 // 45nm: Penryn , Wolfdale, Yorkfield (XE) 588 case 0x1d: // Intel Xeon processor MP. All processors are manufactured using 589 // the 45 nm process. 590 *Type = X86::INTEL_CORE2; // "penryn" 591 *Subtype = X86::INTEL_CORE2_45; 592 break; 593 case 0x1a: // Intel Core i7 processor and Intel Xeon processor. All 594 // processors are manufactured using the 45 nm process. 595 case 0x1e: // Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz. 596 // As found in a Summer 2010 model iMac. 597 case 0x1f: 598 case 0x2e: // Nehalem EX 599 *Type = X86::INTEL_COREI7; // "nehalem" 600 *Subtype = X86::INTEL_COREI7_NEHALEM; 601 break; 602 case 0x25: // Intel Core i7, laptop version. 603 case 0x2c: // Intel Core i7 processor and Intel Xeon processor. All 604 // processors are manufactured using the 32 nm process. 605 case 0x2f: // Westmere EX 606 *Type = X86::INTEL_COREI7; // "westmere" 607 *Subtype = X86::INTEL_COREI7_WESTMERE; 608 break; 609 case 0x2a: // Intel Core i7 processor. All processors are manufactured 610 // using the 32 nm process. 611 case 0x2d: 612 *Type = X86::INTEL_COREI7; //"sandybridge" 613 *Subtype = X86::INTEL_COREI7_SANDYBRIDGE; 614 break; 615 case 0x3a: 616 case 0x3e: // Ivy Bridge EP 617 *Type = X86::INTEL_COREI7; // "ivybridge" 618 *Subtype = X86::INTEL_COREI7_IVYBRIDGE; 619 break; 620 621 // Haswell: 622 case 0x3c: 623 case 0x3f: 624 case 0x45: 625 case 0x46: 626 *Type = X86::INTEL_COREI7; // "haswell" 627 *Subtype = X86::INTEL_COREI7_HASWELL; 628 break; 629 630 // Broadwell: 631 case 0x3d: 632 case 0x47: 633 case 0x4f: 634 case 0x56: 635 *Type = X86::INTEL_COREI7; // "broadwell" 636 *Subtype = X86::INTEL_COREI7_BROADWELL; 637 break; 638 639 // Skylake: 640 case 0x4e: // Skylake mobile 641 case 0x5e: // Skylake desktop 642 case 0x8e: // Kaby Lake mobile 643 case 0x9e: // Kaby Lake desktop 644 *Type = X86::INTEL_COREI7; // "skylake" 645 *Subtype = X86::INTEL_COREI7_SKYLAKE; 646 break; 647 648 // Skylake Xeon: 649 case 0x55: 650 *Type = X86::INTEL_COREI7; 651 *Subtype = X86::INTEL_COREI7_SKYLAKE_AVX512; // "skylake-avx512" 652 break; 653 654 // Cannonlake: 655 case 0x66: 656 *Type = X86::INTEL_COREI7; 657 *Subtype = X86::INTEL_COREI7_CANNONLAKE; // "cannonlake" 658 break; 659 660 case 0x1c: // Most 45 nm Intel Atom processors 661 case 0x26: // 45 nm Atom Lincroft 662 case 0x27: // 32 nm Atom Medfield 663 case 0x35: // 32 nm Atom Midview 664 case 0x36: // 32 nm Atom Midview 665 *Type = X86::INTEL_BONNELL; 666 break; // "bonnell" 667 668 // Atom Silvermont codes from the Intel software optimization guide. 669 case 0x37: 670 case 0x4a: 671 case 0x4d: 672 case 0x5a: 673 case 0x5d: 674 case 0x4c: // really airmont 675 *Type = X86::INTEL_SILVERMONT; 676 break; // "silvermont" 677 // Goldmont: 678 case 0x5c: // Apollo Lake 679 case 0x5f: // Denverton 680 *Type = X86::INTEL_GOLDMONT; 681 break; // "goldmont" 682 case 0x7a: 683 *Type = X86::INTEL_GOLDMONT_PLUS; 684 break; 685 case 0x57: 686 *Type = X86::INTEL_KNL; // knl 687 break; 688 case 0x85: 689 *Type = X86::INTEL_KNM; // knm 690 break; 691 692 default: // Unknown family 6 CPU, try to guess. 693 if (Features & (1 << X86::FEATURE_AVX512VBMI2)) { 694 *Type = X86::INTEL_COREI7; 695 *Subtype = X86::INTEL_COREI7_ICELAKE_CLIENT; 696 break; 697 } 698 699 if (Features & (1 << X86::FEATURE_AVX512VBMI)) { 700 *Type = X86::INTEL_COREI7; 701 *Subtype = X86::INTEL_COREI7_CANNONLAKE; 702 break; 703 } 704 705 if (Features2 & (1 << (X86::FEATURE_AVX512VNNI - 32))) { 706 *Type = X86::INTEL_COREI7; 707 *Subtype = X86::INTEL_COREI7_CASCADELAKE; 708 break; 709 } 710 711 if (Features & (1 << X86::FEATURE_AVX512VL)) { 712 *Type = X86::INTEL_COREI7; 713 *Subtype = X86::INTEL_COREI7_SKYLAKE_AVX512; 714 break; 715 } 716 717 if (Features & (1 << X86::FEATURE_AVX512ER)) { 718 *Type = X86::INTEL_KNL; // knl 719 break; 720 } 721 722 if (Features3 & (1 << (X86::FEATURE_CLFLUSHOPT - 64))) { 723 if (Features3 & (1 << (X86::FEATURE_SHA - 64))) { 724 *Type = X86::INTEL_GOLDMONT; 725 } else { 726 *Type = X86::INTEL_COREI7; 727 *Subtype = X86::INTEL_COREI7_SKYLAKE; 728 } 729 break; 730 } 731 if (Features3 & (1 << (X86::FEATURE_ADX - 64))) { 732 *Type = X86::INTEL_COREI7; 733 *Subtype = X86::INTEL_COREI7_BROADWELL; 734 break; 735 } 736 if (Features & (1 << X86::FEATURE_AVX2)) { 737 *Type = X86::INTEL_COREI7; 738 *Subtype = X86::INTEL_COREI7_HASWELL; 739 break; 740 } 741 if (Features & (1 << X86::FEATURE_AVX)) { 742 *Type = X86::INTEL_COREI7; 743 *Subtype = X86::INTEL_COREI7_SANDYBRIDGE; 744 break; 745 } 746 if (Features & (1 << X86::FEATURE_SSE4_2)) { 747 if (Features3 & (1 << (X86::FEATURE_MOVBE - 64))) { 748 *Type = X86::INTEL_SILVERMONT; 749 } else { 750 *Type = X86::INTEL_COREI7; 751 *Subtype = X86::INTEL_COREI7_NEHALEM; 752 } 753 break; 754 } 755 if (Features & (1 << X86::FEATURE_SSE4_1)) { 756 *Type = X86::INTEL_CORE2; // "penryn" 757 *Subtype = X86::INTEL_CORE2_45; 758 break; 759 } 760 if (Features & (1 << X86::FEATURE_SSSE3)) { 761 if (Features3 & (1 << (X86::FEATURE_MOVBE - 64))) { 762 *Type = X86::INTEL_BONNELL; // "bonnell" 763 } else { 764 *Type = X86::INTEL_CORE2; // "core2" 765 *Subtype = X86::INTEL_CORE2_65; 766 } 767 break; 768 } 769 if (Features3 & (1 << (X86::FEATURE_EM64T - 64))) { 770 *Type = X86::INTEL_CORE2; // "core2" 771 *Subtype = X86::INTEL_CORE2_65; 772 break; 773 } 774 if (Features & (1 << X86::FEATURE_SSE3)) { 775 *Type = X86::INTEL_CORE_DUO; 776 break; 777 } 778 if (Features & (1 << X86::FEATURE_SSE2)) { 779 *Type = X86::INTEL_PENTIUM_M; 780 break; 781 } 782 if (Features & (1 << X86::FEATURE_SSE)) { 783 *Type = X86::INTEL_PENTIUM_III; 784 break; 785 } 786 if (Features & (1 << X86::FEATURE_MMX)) { 787 *Type = X86::INTEL_PENTIUM_II; 788 break; 789 } 790 *Type = X86::INTEL_PENTIUM_PRO; 791 break; 792 } 793 break; 794 case 15: { 795 if (Features3 & (1 << (X86::FEATURE_EM64T - 64))) { 796 *Type = X86::INTEL_NOCONA; 797 break; 798 } 799 if (Features & (1 << X86::FEATURE_SSE3)) { 800 *Type = X86::INTEL_PRESCOTT; 801 break; 802 } 803 *Type = X86::INTEL_PENTIUM_IV; 804 break; 805 } 806 default: 807 break; /*"generic"*/ 808 } 809 } 810 811 static void getAMDProcessorTypeAndSubtype(unsigned Family, unsigned Model, 812 unsigned Features, unsigned *Type, 813 unsigned *Subtype) { 814 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There 815 // appears to be no way to generate the wide variety of AMD-specific targets 816 // from the information returned from CPUID. 817 switch (Family) { 818 case 4: 819 *Type = X86::AMD_i486; 820 break; 821 case 5: 822 *Type = X86::AMDPENTIUM; 823 switch (Model) { 824 case 6: 825 case 7: 826 *Subtype = X86::AMDPENTIUM_K6; 827 break; // "k6" 828 case 8: 829 *Subtype = X86::AMDPENTIUM_K62; 830 break; // "k6-2" 831 case 9: 832 case 13: 833 *Subtype = X86::AMDPENTIUM_K63; 834 break; // "k6-3" 835 case 10: 836 *Subtype = X86::AMDPENTIUM_GEODE; 837 break; // "geode" 838 } 839 break; 840 case 6: 841 if (Features & (1 << X86::FEATURE_SSE)) { 842 *Type = X86::AMD_ATHLON_XP; 843 break; // "athlon-xp" 844 } 845 *Type = X86::AMD_ATHLON; 846 break; // "athlon" 847 case 15: 848 if (Features & (1 << X86::FEATURE_SSE3)) { 849 *Type = X86::AMD_K8SSE3; 850 break; // "k8-sse3" 851 } 852 *Type = X86::AMD_K8; 853 break; // "k8" 854 case 16: 855 *Type = X86::AMDFAM10H; // "amdfam10" 856 switch (Model) { 857 case 2: 858 *Subtype = X86::AMDFAM10H_BARCELONA; 859 break; 860 case 4: 861 *Subtype = X86::AMDFAM10H_SHANGHAI; 862 break; 863 case 8: 864 *Subtype = X86::AMDFAM10H_ISTANBUL; 865 break; 866 } 867 break; 868 case 20: 869 *Type = X86::AMD_BTVER1; 870 break; // "btver1"; 871 case 21: 872 *Type = X86::AMDFAM15H; 873 if (Model >= 0x60 && Model <= 0x7f) { 874 *Subtype = X86::AMDFAM15H_BDVER4; 875 break; // "bdver4"; 60h-7Fh: Excavator 876 } 877 if (Model >= 0x30 && Model <= 0x3f) { 878 *Subtype = X86::AMDFAM15H_BDVER3; 879 break; // "bdver3"; 30h-3Fh: Steamroller 880 } 881 if ((Model >= 0x10 && Model <= 0x1f) || Model == 0x02) { 882 *Subtype = X86::AMDFAM15H_BDVER2; 883 break; // "bdver2"; 02h, 10h-1Fh: Piledriver 884 } 885 if (Model <= 0x0f) { 886 *Subtype = X86::AMDFAM15H_BDVER1; 887 break; // "bdver1"; 00h-0Fh: Bulldozer 888 } 889 break; 890 case 22: 891 *Type = X86::AMD_BTVER2; 892 break; // "btver2" 893 case 23: 894 *Type = X86::AMDFAM17H; 895 *Subtype = X86::AMDFAM17H_ZNVER1; 896 break; 897 default: 898 break; // "generic" 899 } 900 } 901 902 static void getAvailableFeatures(unsigned ECX, unsigned EDX, unsigned MaxLeaf, 903 unsigned *FeaturesOut, unsigned *Features2Out, 904 unsigned *Features3Out) { 905 unsigned Features = 0; 906 unsigned Features2 = 0; 907 unsigned Features3 = 0; 908 unsigned EAX, EBX; 909 910 auto setFeature = [&](unsigned F) { 911 if (F < 32) 912 Features |= 1U << (F & 0x1f); 913 else if (F < 64) 914 Features2 |= 1U << ((F - 32) & 0x1f); 915 else if (F < 96) 916 Features3 |= 1U << ((F - 64) & 0x1f); 917 else 918 llvm_unreachable("Unexpected FeatureBit"); 919 }; 920 921 if ((EDX >> 15) & 1) 922 setFeature(X86::FEATURE_CMOV); 923 if ((EDX >> 23) & 1) 924 setFeature(X86::FEATURE_MMX); 925 if ((EDX >> 25) & 1) 926 setFeature(X86::FEATURE_SSE); 927 if ((EDX >> 26) & 1) 928 setFeature(X86::FEATURE_SSE2); 929 930 if ((ECX >> 0) & 1) 931 setFeature(X86::FEATURE_SSE3); 932 if ((ECX >> 1) & 1) 933 setFeature(X86::FEATURE_PCLMUL); 934 if ((ECX >> 9) & 1) 935 setFeature(X86::FEATURE_SSSE3); 936 if ((ECX >> 12) & 1) 937 setFeature(X86::FEATURE_FMA); 938 if ((ECX >> 19) & 1) 939 setFeature(X86::FEATURE_SSE4_1); 940 if ((ECX >> 20) & 1) 941 setFeature(X86::FEATURE_SSE4_2); 942 if ((ECX >> 23) & 1) 943 setFeature(X86::FEATURE_POPCNT); 944 if ((ECX >> 25) & 1) 945 setFeature(X86::FEATURE_AES); 946 947 if ((ECX >> 22) & 1) 948 setFeature(X86::FEATURE_MOVBE); 949 950 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV 951 // indicates that the AVX registers will be saved and restored on context 952 // switch, then we have full AVX support. 953 const unsigned AVXBits = (1 << 27) | (1 << 28); 954 bool HasAVX = ((ECX & AVXBits) == AVXBits) && !getX86XCR0(&EAX, &EDX) && 955 ((EAX & 0x6) == 0x6); 956 bool HasAVX512Save = HasAVX && ((EAX & 0xe0) == 0xe0); 957 958 if (HasAVX) 959 setFeature(X86::FEATURE_AVX); 960 961 bool HasLeaf7 = 962 MaxLeaf >= 0x7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX); 963 964 if (HasLeaf7 && ((EBX >> 3) & 1)) 965 setFeature(X86::FEATURE_BMI); 966 if (HasLeaf7 && ((EBX >> 5) & 1) && HasAVX) 967 setFeature(X86::FEATURE_AVX2); 968 if (HasLeaf7 && ((EBX >> 9) & 1)) 969 setFeature(X86::FEATURE_BMI2); 970 if (HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save) 971 setFeature(X86::FEATURE_AVX512F); 972 if (HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save) 973 setFeature(X86::FEATURE_AVX512DQ); 974 if (HasLeaf7 && ((EBX >> 19) & 1)) 975 setFeature(X86::FEATURE_ADX); 976 if (HasLeaf7 && ((EBX >> 21) & 1) && HasAVX512Save) 977 setFeature(X86::FEATURE_AVX512IFMA); 978 if (HasLeaf7 && ((EBX >> 23) & 1)) 979 setFeature(X86::FEATURE_CLFLUSHOPT); 980 if (HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save) 981 setFeature(X86::FEATURE_AVX512PF); 982 if (HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save) 983 setFeature(X86::FEATURE_AVX512ER); 984 if (HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save) 985 setFeature(X86::FEATURE_AVX512CD); 986 if (HasLeaf7 && ((EBX >> 29) & 1)) 987 setFeature(X86::FEATURE_SHA); 988 if (HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save) 989 setFeature(X86::FEATURE_AVX512BW); 990 if (HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save) 991 setFeature(X86::FEATURE_AVX512VL); 992 993 if (HasLeaf7 && ((ECX >> 1) & 1) && HasAVX512Save) 994 setFeature(X86::FEATURE_AVX512VBMI); 995 if (HasLeaf7 && ((ECX >> 6) & 1) && HasAVX512Save) 996 setFeature(X86::FEATURE_AVX512VBMI2); 997 if (HasLeaf7 && ((ECX >> 8) & 1)) 998 setFeature(X86::FEATURE_GFNI); 999 if (HasLeaf7 && ((ECX >> 10) & 1) && HasAVX) 1000 setFeature(X86::FEATURE_VPCLMULQDQ); 1001 if (HasLeaf7 && ((ECX >> 11) & 1) && HasAVX512Save) 1002 setFeature(X86::FEATURE_AVX512VNNI); 1003 if (HasLeaf7 && ((ECX >> 12) & 1) && HasAVX512Save) 1004 setFeature(X86::FEATURE_AVX512BITALG); 1005 if (HasLeaf7 && ((ECX >> 14) & 1) && HasAVX512Save) 1006 setFeature(X86::FEATURE_AVX512VPOPCNTDQ); 1007 1008 if (HasLeaf7 && ((EDX >> 2) & 1) && HasAVX512Save) 1009 setFeature(X86::FEATURE_AVX5124VNNIW); 1010 if (HasLeaf7 && ((EDX >> 3) & 1) && HasAVX512Save) 1011 setFeature(X86::FEATURE_AVX5124FMAPS); 1012 1013 unsigned MaxExtLevel; 1014 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX); 1015 1016 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 && 1017 !getX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX); 1018 if (HasExtLeaf1 && ((ECX >> 6) & 1)) 1019 setFeature(X86::FEATURE_SSE4_A); 1020 if (HasExtLeaf1 && ((ECX >> 11) & 1)) 1021 setFeature(X86::FEATURE_XOP); 1022 if (HasExtLeaf1 && ((ECX >> 16) & 1)) 1023 setFeature(X86::FEATURE_FMA4); 1024 1025 if (HasExtLeaf1 && ((EDX >> 29) & 1)) 1026 setFeature(X86::FEATURE_EM64T); 1027 1028 *FeaturesOut = Features; 1029 *Features2Out = Features2; 1030 *Features3Out = Features3; 1031 } 1032 1033 StringRef sys::getHostCPUName() { 1034 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0; 1035 unsigned MaxLeaf, Vendor; 1036 1037 #if defined(__GNUC__) || defined(__clang__) 1038 //FIXME: include cpuid.h from clang or copy __get_cpuid_max here 1039 // and simplify it to not invoke __cpuid (like cpu_model.c in 1040 // compiler-rt/lib/builtins/cpu_model.c? 1041 // Opting for the second option. 1042 if(!isCpuIdSupported()) 1043 return "generic"; 1044 #endif 1045 if (getX86CpuIDAndInfo(0, &MaxLeaf, &Vendor, &ECX, &EDX) || MaxLeaf < 1) 1046 return "generic"; 1047 getX86CpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX); 1048 1049 unsigned Brand_id = EBX & 0xff; 1050 unsigned Family = 0, Model = 0; 1051 unsigned Features = 0, Features2 = 0, Features3 = 0; 1052 detectX86FamilyModel(EAX, &Family, &Model); 1053 getAvailableFeatures(ECX, EDX, MaxLeaf, &Features, &Features2, &Features3); 1054 1055 unsigned Type = 0; 1056 unsigned Subtype = 0; 1057 1058 if (Vendor == SIG_INTEL) { 1059 getIntelProcessorTypeAndSubtype(Family, Model, Brand_id, Features, 1060 Features2, Features3, &Type, &Subtype); 1061 } else if (Vendor == SIG_AMD) { 1062 getAMDProcessorTypeAndSubtype(Family, Model, Features, &Type, &Subtype); 1063 } 1064 1065 // Check subtypes first since those are more specific. 1066 #define X86_CPU_SUBTYPE(ARCHNAME, ENUM) \ 1067 if (Subtype == X86::ENUM) \ 1068 return ARCHNAME; 1069 #include "llvm/Support/X86TargetParser.def" 1070 1071 // Now check types. 1072 #define X86_CPU_TYPE(ARCHNAME, ENUM) \ 1073 if (Type == X86::ENUM) \ 1074 return ARCHNAME; 1075 #include "llvm/Support/X86TargetParser.def" 1076 1077 return "generic"; 1078 } 1079 1080 #elif defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__)) 1081 StringRef sys::getHostCPUName() { 1082 host_basic_info_data_t hostInfo; 1083 mach_msg_type_number_t infoCount; 1084 1085 infoCount = HOST_BASIC_INFO_COUNT; 1086 mach_port_t hostPort = mach_host_self(); 1087 host_info(hostPort, HOST_BASIC_INFO, (host_info_t)&hostInfo, 1088 &infoCount); 1089 mach_port_deallocate(mach_task_self(), hostPort); 1090 1091 if (hostInfo.cpu_type != CPU_TYPE_POWERPC) 1092 return "generic"; 1093 1094 switch (hostInfo.cpu_subtype) { 1095 case CPU_SUBTYPE_POWERPC_601: 1096 return "601"; 1097 case CPU_SUBTYPE_POWERPC_602: 1098 return "602"; 1099 case CPU_SUBTYPE_POWERPC_603: 1100 return "603"; 1101 case CPU_SUBTYPE_POWERPC_603e: 1102 return "603e"; 1103 case CPU_SUBTYPE_POWERPC_603ev: 1104 return "603ev"; 1105 case CPU_SUBTYPE_POWERPC_604: 1106 return "604"; 1107 case CPU_SUBTYPE_POWERPC_604e: 1108 return "604e"; 1109 case CPU_SUBTYPE_POWERPC_620: 1110 return "620"; 1111 case CPU_SUBTYPE_POWERPC_750: 1112 return "750"; 1113 case CPU_SUBTYPE_POWERPC_7400: 1114 return "7400"; 1115 case CPU_SUBTYPE_POWERPC_7450: 1116 return "7450"; 1117 case CPU_SUBTYPE_POWERPC_970: 1118 return "970"; 1119 default:; 1120 } 1121 1122 return "generic"; 1123 } 1124 #elif defined(__linux__) && (defined(__ppc__) || defined(__powerpc__)) 1125 StringRef sys::getHostCPUName() { 1126 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent(); 1127 StringRef Content = P ? P->getBuffer() : ""; 1128 return detail::getHostCPUNameForPowerPC(Content); 1129 } 1130 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__)) 1131 StringRef sys::getHostCPUName() { 1132 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent(); 1133 StringRef Content = P ? P->getBuffer() : ""; 1134 return detail::getHostCPUNameForARM(Content); 1135 } 1136 #elif defined(__linux__) && defined(__s390x__) 1137 StringRef sys::getHostCPUName() { 1138 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent(); 1139 StringRef Content = P ? P->getBuffer() : ""; 1140 return detail::getHostCPUNameForS390x(Content); 1141 } 1142 #else 1143 StringRef sys::getHostCPUName() { return "generic"; } 1144 #endif 1145 1146 #if defined(__linux__) && defined(__x86_64__) 1147 // On Linux, the number of physical cores can be computed from /proc/cpuinfo, 1148 // using the number of unique physical/core id pairs. The following 1149 // implementation reads the /proc/cpuinfo format on an x86_64 system. 1150 static int computeHostNumPhysicalCores() { 1151 // Read /proc/cpuinfo as a stream (until EOF reached). It cannot be 1152 // mmapped because it appears to have 0 size. 1153 llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> Text = 1154 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo"); 1155 if (std::error_code EC = Text.getError()) { 1156 llvm::errs() << "Can't read " 1157 << "/proc/cpuinfo: " << EC.message() << "\n"; 1158 return -1; 1159 } 1160 SmallVector<StringRef, 8> strs; 1161 (*Text)->getBuffer().split(strs, "\n", /*MaxSplit=*/-1, 1162 /*KeepEmpty=*/false); 1163 int CurPhysicalId = -1; 1164 int CurCoreId = -1; 1165 SmallSet<std::pair<int, int>, 32> UniqueItems; 1166 for (auto &Line : strs) { 1167 Line = Line.trim(); 1168 if (!Line.startswith("physical id") && !Line.startswith("core id")) 1169 continue; 1170 std::pair<StringRef, StringRef> Data = Line.split(':'); 1171 auto Name = Data.first.trim(); 1172 auto Val = Data.second.trim(); 1173 if (Name == "physical id") { 1174 assert(CurPhysicalId == -1 && 1175 "Expected a core id before seeing another physical id"); 1176 Val.getAsInteger(10, CurPhysicalId); 1177 } 1178 if (Name == "core id") { 1179 assert(CurCoreId == -1 && 1180 "Expected a physical id before seeing another core id"); 1181 Val.getAsInteger(10, CurCoreId); 1182 } 1183 if (CurPhysicalId != -1 && CurCoreId != -1) { 1184 UniqueItems.insert(std::make_pair(CurPhysicalId, CurCoreId)); 1185 CurPhysicalId = -1; 1186 CurCoreId = -1; 1187 } 1188 } 1189 return UniqueItems.size(); 1190 } 1191 #elif defined(__APPLE__) && defined(__x86_64__) 1192 #include <sys/param.h> 1193 #include <sys/sysctl.h> 1194 1195 // Gets the number of *physical cores* on the machine. 1196 static int computeHostNumPhysicalCores() { 1197 uint32_t count; 1198 size_t len = sizeof(count); 1199 sysctlbyname("hw.physicalcpu", &count, &len, NULL, 0); 1200 if (count < 1) { 1201 int nm[2]; 1202 nm[0] = CTL_HW; 1203 nm[1] = HW_AVAILCPU; 1204 sysctl(nm, 2, &count, &len, NULL, 0); 1205 if (count < 1) 1206 return -1; 1207 } 1208 return count; 1209 } 1210 #else 1211 // On other systems, return -1 to indicate unknown. 1212 static int computeHostNumPhysicalCores() { return -1; } 1213 #endif 1214 1215 int sys::getHostNumPhysicalCores() { 1216 static int NumCores = computeHostNumPhysicalCores(); 1217 return NumCores; 1218 } 1219 1220 #if defined(__i386__) || defined(_M_IX86) || \ 1221 defined(__x86_64__) || defined(_M_X64) 1222 bool sys::getHostCPUFeatures(StringMap<bool> &Features) { 1223 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0; 1224 unsigned MaxLevel; 1225 union { 1226 unsigned u[3]; 1227 char c[12]; 1228 } text; 1229 1230 if (getX86CpuIDAndInfo(0, &MaxLevel, text.u + 0, text.u + 2, text.u + 1) || 1231 MaxLevel < 1) 1232 return false; 1233 1234 getX86CpuIDAndInfo(1, &EAX, &EBX, &ECX, &EDX); 1235 1236 Features["cmov"] = (EDX >> 15) & 1; 1237 Features["mmx"] = (EDX >> 23) & 1; 1238 Features["sse"] = (EDX >> 25) & 1; 1239 Features["sse2"] = (EDX >> 26) & 1; 1240 1241 Features["sse3"] = (ECX >> 0) & 1; 1242 Features["pclmul"] = (ECX >> 1) & 1; 1243 Features["ssse3"] = (ECX >> 9) & 1; 1244 Features["cx16"] = (ECX >> 13) & 1; 1245 Features["sse4.1"] = (ECX >> 19) & 1; 1246 Features["sse4.2"] = (ECX >> 20) & 1; 1247 Features["movbe"] = (ECX >> 22) & 1; 1248 Features["popcnt"] = (ECX >> 23) & 1; 1249 Features["aes"] = (ECX >> 25) & 1; 1250 Features["rdrnd"] = (ECX >> 30) & 1; 1251 1252 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV 1253 // indicates that the AVX registers will be saved and restored on context 1254 // switch, then we have full AVX support. 1255 bool HasAVXSave = ((ECX >> 27) & 1) && ((ECX >> 28) & 1) && 1256 !getX86XCR0(&EAX, &EDX) && ((EAX & 0x6) == 0x6); 1257 // AVX512 requires additional context to be saved by the OS. 1258 bool HasAVX512Save = HasAVXSave && ((EAX & 0xe0) == 0xe0); 1259 1260 Features["avx"] = HasAVXSave; 1261 Features["fma"] = ((ECX >> 12) & 1) && HasAVXSave; 1262 // Only enable XSAVE if OS has enabled support for saving YMM state. 1263 Features["xsave"] = ((ECX >> 26) & 1) && HasAVXSave; 1264 Features["f16c"] = ((ECX >> 29) & 1) && HasAVXSave; 1265 1266 unsigned MaxExtLevel; 1267 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX); 1268 1269 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 && 1270 !getX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX); 1271 Features["sahf"] = HasExtLeaf1 && ((ECX >> 0) & 1); 1272 Features["lzcnt"] = HasExtLeaf1 && ((ECX >> 5) & 1); 1273 Features["sse4a"] = HasExtLeaf1 && ((ECX >> 6) & 1); 1274 Features["prfchw"] = HasExtLeaf1 && ((ECX >> 8) & 1); 1275 Features["xop"] = HasExtLeaf1 && ((ECX >> 11) & 1) && HasAVXSave; 1276 Features["lwp"] = HasExtLeaf1 && ((ECX >> 15) & 1); 1277 Features["fma4"] = HasExtLeaf1 && ((ECX >> 16) & 1) && HasAVXSave; 1278 Features["tbm"] = HasExtLeaf1 && ((ECX >> 21) & 1); 1279 Features["mwaitx"] = HasExtLeaf1 && ((ECX >> 29) & 1); 1280 1281 Features["64bit"] = HasExtLeaf1 && ((EDX >> 29) & 1); 1282 1283 // Miscellaneous memory related features, detected by 1284 // using the 0x80000008 leaf of the CPUID instruction 1285 bool HasExtLeaf8 = MaxExtLevel >= 0x80000008 && 1286 !getX86CpuIDAndInfo(0x80000008, &EAX, &EBX, &ECX, &EDX); 1287 Features["clzero"] = HasExtLeaf8 && ((EBX >> 0) & 1); 1288 Features["wbnoinvd"] = HasExtLeaf8 && ((EBX >> 9) & 1); 1289 1290 bool HasLeaf7 = 1291 MaxLevel >= 7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX); 1292 1293 Features["fsgsbase"] = HasLeaf7 && ((EBX >> 0) & 1); 1294 Features["sgx"] = HasLeaf7 && ((EBX >> 2) & 1); 1295 Features["bmi"] = HasLeaf7 && ((EBX >> 3) & 1); 1296 // AVX2 is only supported if we have the OS save support from AVX. 1297 Features["avx2"] = HasLeaf7 && ((EBX >> 5) & 1) && HasAVXSave; 1298 Features["bmi2"] = HasLeaf7 && ((EBX >> 8) & 1); 1299 Features["invpcid"] = HasLeaf7 && ((EBX >> 10) & 1); 1300 Features["rtm"] = HasLeaf7 && ((EBX >> 11) & 1); 1301 // AVX512 is only supported if the OS supports the context save for it. 1302 Features["avx512f"] = HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save; 1303 Features["avx512dq"] = HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save; 1304 Features["rdseed"] = HasLeaf7 && ((EBX >> 18) & 1); 1305 Features["adx"] = HasLeaf7 && ((EBX >> 19) & 1); 1306 Features["avx512ifma"] = HasLeaf7 && ((EBX >> 21) & 1) && HasAVX512Save; 1307 Features["clflushopt"] = HasLeaf7 && ((EBX >> 23) & 1); 1308 Features["clwb"] = HasLeaf7 && ((EBX >> 24) & 1); 1309 Features["avx512pf"] = HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save; 1310 Features["avx512er"] = HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save; 1311 Features["avx512cd"] = HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save; 1312 Features["sha"] = HasLeaf7 && ((EBX >> 29) & 1); 1313 Features["avx512bw"] = HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save; 1314 Features["avx512vl"] = HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save; 1315 1316 Features["prefetchwt1"] = HasLeaf7 && ((ECX >> 0) & 1); 1317 Features["avx512vbmi"] = HasLeaf7 && ((ECX >> 1) & 1) && HasAVX512Save; 1318 Features["pku"] = HasLeaf7 && ((ECX >> 4) & 1); 1319 Features["waitpkg"] = HasLeaf7 && ((ECX >> 5) & 1); 1320 Features["avx512vbmi2"] = HasLeaf7 && ((ECX >> 6) & 1) && HasAVX512Save; 1321 Features["shstk"] = HasLeaf7 && ((ECX >> 7) & 1); 1322 Features["gfni"] = HasLeaf7 && ((ECX >> 8) & 1); 1323 Features["vaes"] = HasLeaf7 && ((ECX >> 9) & 1) && HasAVXSave; 1324 Features["vpclmulqdq"] = HasLeaf7 && ((ECX >> 10) & 1) && HasAVXSave; 1325 Features["avx512vnni"] = HasLeaf7 && ((ECX >> 11) & 1) && HasAVX512Save; 1326 Features["avx512bitalg"] = HasLeaf7 && ((ECX >> 12) & 1) && HasAVX512Save; 1327 Features["avx512vpopcntdq"] = HasLeaf7 && ((ECX >> 14) & 1) && HasAVX512Save; 1328 Features["rdpid"] = HasLeaf7 && ((ECX >> 22) & 1); 1329 Features["cldemote"] = HasLeaf7 && ((ECX >> 25) & 1); 1330 Features["movdiri"] = HasLeaf7 && ((ECX >> 27) & 1); 1331 Features["movdir64b"] = HasLeaf7 && ((ECX >> 28) & 1); 1332 1333 // There are two CPUID leafs which information associated with the pconfig 1334 // instruction: 1335 // EAX=0x7, ECX=0x0 indicates the availability of the instruction (via the 18th 1336 // bit of EDX), while the EAX=0x1b leaf returns information on the 1337 // availability of specific pconfig leafs. 1338 // The target feature here only refers to the the first of these two. 1339 // Users might need to check for the availability of specific pconfig 1340 // leaves using cpuid, since that information is ignored while 1341 // detecting features using the "-march=native" flag. 1342 // For more info, see X86 ISA docs. 1343 Features["pconfig"] = HasLeaf7 && ((EDX >> 18) & 1); 1344 1345 bool HasLeafD = MaxLevel >= 0xd && 1346 !getX86CpuIDAndInfoEx(0xd, 0x1, &EAX, &EBX, &ECX, &EDX); 1347 1348 // Only enable XSAVE if OS has enabled support for saving YMM state. 1349 Features["xsaveopt"] = HasLeafD && ((EAX >> 0) & 1) && HasAVXSave; 1350 Features["xsavec"] = HasLeafD && ((EAX >> 1) & 1) && HasAVXSave; 1351 Features["xsaves"] = HasLeafD && ((EAX >> 3) & 1) && HasAVXSave; 1352 1353 bool HasLeaf14 = MaxLevel >= 0x14 && 1354 !getX86CpuIDAndInfoEx(0x14, 0x0, &EAX, &EBX, &ECX, &EDX); 1355 1356 Features["ptwrite"] = HasLeaf14 && ((EBX >> 4) & 1); 1357 1358 return true; 1359 } 1360 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__)) 1361 bool sys::getHostCPUFeatures(StringMap<bool> &Features) { 1362 std::unique_ptr<llvm::MemoryBuffer> P = getProcCpuinfoContent(); 1363 if (!P) 1364 return false; 1365 1366 SmallVector<StringRef, 32> Lines; 1367 P->getBuffer().split(Lines, "\n"); 1368 1369 SmallVector<StringRef, 32> CPUFeatures; 1370 1371 // Look for the CPU features. 1372 for (unsigned I = 0, E = Lines.size(); I != E; ++I) 1373 if (Lines[I].startswith("Features")) { 1374 Lines[I].split(CPUFeatures, ' '); 1375 break; 1376 } 1377 1378 #if defined(__aarch64__) 1379 // Keep track of which crypto features we have seen 1380 enum { CAP_AES = 0x1, CAP_PMULL = 0x2, CAP_SHA1 = 0x4, CAP_SHA2 = 0x8 }; 1381 uint32_t crypto = 0; 1382 #endif 1383 1384 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) { 1385 StringRef LLVMFeatureStr = StringSwitch<StringRef>(CPUFeatures[I]) 1386 #if defined(__aarch64__) 1387 .Case("asimd", "neon") 1388 .Case("fp", "fp-armv8") 1389 .Case("crc32", "crc") 1390 #else 1391 .Case("half", "fp16") 1392 .Case("neon", "neon") 1393 .Case("vfpv3", "vfp3") 1394 .Case("vfpv3d16", "d16") 1395 .Case("vfpv4", "vfp4") 1396 .Case("idiva", "hwdiv-arm") 1397 .Case("idivt", "hwdiv") 1398 #endif 1399 .Default(""); 1400 1401 #if defined(__aarch64__) 1402 // We need to check crypto separately since we need all of the crypto 1403 // extensions to enable the subtarget feature 1404 if (CPUFeatures[I] == "aes") 1405 crypto |= CAP_AES; 1406 else if (CPUFeatures[I] == "pmull") 1407 crypto |= CAP_PMULL; 1408 else if (CPUFeatures[I] == "sha1") 1409 crypto |= CAP_SHA1; 1410 else if (CPUFeatures[I] == "sha2") 1411 crypto |= CAP_SHA2; 1412 #endif 1413 1414 if (LLVMFeatureStr != "") 1415 Features[LLVMFeatureStr] = true; 1416 } 1417 1418 #if defined(__aarch64__) 1419 // If we have all crypto bits we can add the feature 1420 if (crypto == (CAP_AES | CAP_PMULL | CAP_SHA1 | CAP_SHA2)) 1421 Features["crypto"] = true; 1422 #endif 1423 1424 return true; 1425 } 1426 #else 1427 bool sys::getHostCPUFeatures(StringMap<bool> &Features) { return false; } 1428 #endif 1429 1430 std::string sys::getProcessTriple() { 1431 std::string TargetTripleString = updateTripleOSVersion(LLVM_HOST_TRIPLE); 1432 Triple PT(Triple::normalize(TargetTripleString)); 1433 1434 if (sizeof(void *) == 8 && PT.isArch32Bit()) 1435 PT = PT.get64BitArchVariant(); 1436 if (sizeof(void *) == 4 && PT.isArch64Bit()) 1437 PT = PT.get32BitArchVariant(); 1438 1439 return PT.str(); 1440 } 1441