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