1 //===-- Host.cpp - Implement OS Host Concept --------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the operating system Host concept.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Support/Host.h"
15 #include "llvm/ADT/SmallVector.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/ADT/StringSwitch.h"
18 #include "llvm/ADT/Triple.h"
19 #include "llvm/Config/config.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Support/FileSystem.h"
22 #include "llvm/Support/raw_ostream.h"
25 // Include the platform-specific parts of this class.
27 #include "Unix/Host.inc"
30 #include "Windows/Host.inc"
35 #if defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
36 #include <mach/mach.h>
37 #include <mach/mach_host.h>
38 #include <mach/host_info.h>
39 #include <mach/machine.h>
42 #define DEBUG_TYPE "host-detection"
44 //===----------------------------------------------------------------------===//
46 // Implementations of the CPU detection routines
48 //===----------------------------------------------------------------------===//
52 #if defined(__linux__)
53 static ssize_t LLVM_ATTRIBUTE_UNUSED readCpuInfo(void *Buf, size_t Size) {
54 // Note: We cannot mmap /proc/cpuinfo here and then process the resulting
55 // memory buffer because the 'file' has 0 size (it can be read from only
59 std::error_code EC = sys::fs::openFileForRead("/proc/cpuinfo", FD);
61 DEBUG(dbgs() << "Unable to open /proc/cpuinfo: " << EC.message() << "\n");
64 int Ret = read(FD, Buf, Size);
65 int CloseStatus = close(FD);
72 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\
73 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
75 /// GetX86CpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
76 /// specified arguments. If we can't run cpuid on the host, return true.
77 static bool GetX86CpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX,
78 unsigned *rECX, unsigned *rEDX) {
79 #if defined(__GNUC__) || defined(__clang__)
80 #if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
81 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
82 asm ("movq\t%%rbx, %%rsi\n\t"
84 "xchgq\t%%rbx, %%rsi\n\t"
91 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
92 asm ("movl\t%%ebx, %%esi\n\t"
94 "xchgl\t%%ebx, %%esi\n\t"
101 // pedantic #else returns to appease -Wunreachable-code (so we don't generate
102 // postprocessed code that looks like "return true; return false;")
106 #elif defined(_MSC_VER)
107 // The MSVC intrinsic is portable across x86 and x64.
109 __cpuid(registers, value);
110 *rEAX = registers[0];
111 *rEBX = registers[1];
112 *rECX = registers[2];
113 *rEDX = registers[3];
120 /// GetX86CpuIDAndInfoEx - Execute the specified cpuid with subleaf and return the
121 /// 4 values in the specified arguments. If we can't run cpuid on the host,
123 static bool GetX86CpuIDAndInfoEx(unsigned value, unsigned subleaf,
124 unsigned *rEAX, unsigned *rEBX, unsigned *rECX,
126 #if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
127 #if defined(__GNUC__)
128 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
129 asm ("movq\t%%rbx, %%rsi\n\t"
131 "xchgq\t%%rbx, %%rsi\n\t"
139 #elif defined(_MSC_VER)
141 __cpuidex(registers, value, subleaf);
142 *rEAX = registers[0];
143 *rEBX = registers[1];
144 *rECX = registers[2];
145 *rEDX = registers[3];
150 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
151 #if defined(__GNUC__)
152 asm ("movl\t%%ebx, %%esi\n\t"
154 "xchgl\t%%ebx, %%esi\n\t"
162 #elif defined(_MSC_VER)
168 mov dword ptr [esi],eax
170 mov dword ptr [esi],ebx
172 mov dword ptr [esi],ecx
174 mov dword ptr [esi],edx
185 static bool GetX86XCR0(unsigned *rEAX, unsigned *rEDX) {
186 #if defined(__GNUC__)
187 // Check xgetbv; this uses a .byte sequence instead of the instruction
188 // directly because older assemblers do not include support for xgetbv and
189 // there is no easy way to conditionally compile based on the assembler used.
190 __asm__ (".byte 0x0f, 0x01, 0xd0" : "=a" (*rEAX), "=d" (*rEDX) : "c" (0));
192 #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
193 unsigned long long Result = _xgetbv(_XCR_XFEATURE_ENABLED_MASK);
195 *rEDX = Result >> 32;
202 static void DetectX86FamilyModel(unsigned EAX, unsigned &Family,
204 Family = (EAX >> 8) & 0xf; // Bits 8 - 11
205 Model = (EAX >> 4) & 0xf; // Bits 4 - 7
206 if (Family == 6 || Family == 0xf) {
208 // Examine extended family ID if family ID is F.
209 Family += (EAX >> 20) & 0xff; // Bits 20 - 27
210 // Examine extended model ID if family ID is 6 or F.
211 Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19
215 StringRef sys::getHostCPUName() {
216 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
217 if (GetX86CpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX))
221 DetectX86FamilyModel(EAX, Family, Model);
228 GetX86CpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1);
230 unsigned MaxLeaf = EAX;
231 bool HasSSE3 = (ECX & 0x1);
232 bool HasSSE41 = (ECX & 0x80000);
233 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
234 // indicates that the AVX registers will be saved and restored on context
235 // switch, then we have full AVX support.
236 const unsigned AVXBits = (1 << 27) | (1 << 28);
237 bool HasAVX = ((ECX & AVXBits) == AVXBits) && !GetX86XCR0(&EAX, &EDX) &&
238 ((EAX & 0x6) == 0x6);
239 bool HasAVX2 = HasAVX && MaxLeaf >= 0x7 &&
240 !GetX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX) &&
242 GetX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
243 bool Em64T = (EDX >> 29) & 0x1;
244 bool HasTBM = (ECX >> 21) & 0x1;
246 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
252 case 0: // Intel486 DX processors
253 case 1: // Intel486 DX processors
254 case 2: // Intel486 SX processors
255 case 3: // Intel487 processors, IntelDX2 OverDrive processors,
256 // IntelDX2 processors
257 case 4: // Intel486 SL processor
258 case 5: // IntelSX2 processors
259 case 7: // Write-Back Enhanced IntelDX2 processors
260 case 8: // IntelDX4 OverDrive processors, IntelDX4 processors
261 default: return "i486";
265 case 1: // Pentium OverDrive processor for Pentium processor (60, 66),
266 // Pentium processors (60, 66)
267 case 2: // Pentium OverDrive processor for Pentium processor (75, 90,
268 // 100, 120, 133), Pentium processors (75, 90, 100, 120, 133,
270 case 3: // Pentium OverDrive processors for Intel486 processor-based
274 case 4: // Pentium OverDrive processor with MMX technology for Pentium
275 // processor (75, 90, 100, 120, 133), Pentium processor with
276 // MMX technology (166, 200)
277 return "pentium-mmx";
279 default: return "pentium";
283 case 1: // Pentium Pro processor
286 case 3: // Intel Pentium II OverDrive processor, Pentium II processor,
288 case 5: // Pentium II processor, model 05, Pentium II Xeon processor,
289 // model 05, and Intel Celeron processor, model 05
290 case 6: // Celeron processor, model 06
293 case 7: // Pentium III processor, model 07, and Pentium III Xeon
294 // processor, model 07
295 case 8: // Pentium III processor, model 08, Pentium III Xeon processor,
296 // model 08, and Celeron processor, model 08
297 case 10: // Pentium III Xeon processor, model 0Ah
298 case 11: // Pentium III processor, model 0Bh
301 case 9: // Intel Pentium M processor, Intel Celeron M processor model 09.
302 case 13: // Intel Pentium M processor, Intel Celeron M processor, model
303 // 0Dh. All processors are manufactured using the 90 nm process.
304 case 21: // Intel EP80579 Integrated Processor and Intel EP80579
305 // Integrated Processor with Intel QuickAssist Technology
308 case 14: // Intel Core Duo processor, Intel Core Solo processor, model
309 // 0Eh. All processors are manufactured using the 65 nm process.
312 case 15: // Intel Core 2 Duo processor, Intel Core 2 Duo mobile
313 // processor, Intel Core 2 Quad processor, Intel Core 2 Quad
314 // mobile processor, Intel Core 2 Extreme processor, Intel
315 // Pentium Dual-Core processor, Intel Xeon processor, model
316 // 0Fh. All processors are manufactured using the 65 nm process.
317 case 22: // Intel Celeron processor model 16h. All processors are
318 // manufactured using the 65 nm process
321 case 21: // Intel EP80579 Integrated Processor and Intel EP80579
322 // Integrated Processor with Intel QuickAssist Technology
323 return "i686"; // FIXME: ???
325 case 23: // Intel Core 2 Extreme processor, Intel Xeon processor, model
326 // 17h. All processors are manufactured using the 45 nm process.
328 // 45nm: Penryn , Wolfdale, Yorkfield (XE)
329 case 29: // Intel Xeon processor MP. All processors are manufactured using
330 // the 45 nm process.
331 // Not all Penryn processors support SSE 4.1 (such as the Pentium brand)
332 return HasSSE41 ? "penryn" : "core2";
334 case 26: // Intel Core i7 processor and Intel Xeon processor. All
335 // processors are manufactured using the 45 nm process.
336 case 30: // Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz.
337 // As found in a Summer 2010 model iMac.
338 case 37: // Intel Core i7, laptop version.
339 case 44: // Intel Core i7 processor and Intel Xeon processor. All
340 // processors are manufactured using the 32 nm process.
341 case 46: // Nehalem EX
342 case 47: // Westmere EX
346 case 42: // Intel Core i7 processor. All processors are manufactured
347 // using the 32 nm process.
349 // Not all Sandy Bridge processors support AVX (such as the Pentium
350 // versions instead of the i7 versions).
351 return HasAVX ? "corei7-avx" : "corei7";
355 case 62: // Ivy Bridge EP
356 // Not all Ivy Bridge processors support AVX (such as the Pentium
357 // versions instead of the i7 versions).
358 return HasAVX ? "core-avx-i" : "corei7";
365 // Not all Haswell processors support AVX2 (such as the Pentium
366 // versions instead of the i7 versions).
367 return HasAVX2 ? "core-avx2" : "corei7";
371 // Not all Broadwell processors support AVX2 (such as the Pentium
372 // versions instead of the i7 versions).
373 return HasAVX2 ? "broadwell" : "corei7";
375 case 28: // Most 45 nm Intel Atom processors
376 case 38: // 45 nm Atom Lincroft
377 case 39: // 32 nm Atom Medfield
378 case 53: // 32 nm Atom Midview
379 case 54: // 32 nm Atom Midview
382 // Atom Silvermont codes from the Intel software optimization guide.
388 default: return (Em64T) ? "x86-64" : "i686";
392 case 0: // Pentium 4 processor, Intel Xeon processor. All processors are
393 // model 00h and manufactured using the 0.18 micron process.
394 case 1: // Pentium 4 processor, Intel Xeon processor, Intel Xeon
395 // processor MP, and Intel Celeron processor. All processors are
396 // model 01h and manufactured using the 0.18 micron process.
397 case 2: // Pentium 4 processor, Mobile Intel Pentium 4 processor - M,
398 // Intel Xeon processor, Intel Xeon processor MP, Intel Celeron
399 // processor, and Mobile Intel Celeron processor. All processors
400 // are model 02h and manufactured using the 0.13 micron process.
401 return (Em64T) ? "x86-64" : "pentium4";
403 case 3: // Pentium 4 processor, Intel Xeon processor, Intel Celeron D
404 // processor. All processors are model 03h and manufactured using
405 // the 90 nm process.
406 case 4: // Pentium 4 processor, Pentium 4 processor Extreme Edition,
407 // Pentium D processor, Intel Xeon processor, Intel Xeon
408 // processor MP, Intel Celeron D processor. All processors are
409 // model 04h and manufactured using the 90 nm process.
410 case 6: // Pentium 4 processor, Pentium D processor, Pentium processor
411 // Extreme Edition, Intel Xeon processor, Intel Xeon processor
412 // MP, Intel Celeron D processor. All processors are model 06h
413 // and manufactured using the 65 nm process.
414 return (Em64T) ? "nocona" : "prescott";
417 return (Em64T) ? "x86-64" : "pentium4";
424 } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
425 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
426 // appears to be no way to generate the wide variety of AMD-specific targets
427 // from the information returned from CPUID.
435 case 8: return "k6-2";
437 case 13: return "k6-3";
438 case 10: return "geode";
439 default: return "pentium";
443 case 4: return "athlon-tbird";
446 case 8: return "athlon-mp";
447 case 10: return "athlon-xp";
448 default: return "athlon";
454 case 1: return "opteron";
455 case 5: return "athlon-fx"; // also opteron
456 default: return "athlon64";
463 if (!HasAVX) // If the OS doesn't support AVX provide a sane fallback.
466 return "bdver4"; // 50h-6Fh: Excavator
468 return "bdver3"; // 30h-3Fh: Steamroller
469 if (Model >= 0x10 || HasTBM)
470 return "bdver2"; // 10h-1Fh: Piledriver
471 return "bdver1"; // 00h-0Fh: Bulldozer
473 if (!HasAVX) // If the OS doesn't support AVX provide a sane fallback.
482 #elif defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
483 StringRef sys::getHostCPUName() {
484 host_basic_info_data_t hostInfo;
485 mach_msg_type_number_t infoCount;
487 infoCount = HOST_BASIC_INFO_COUNT;
488 host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&hostInfo,
491 if (hostInfo.cpu_type != CPU_TYPE_POWERPC) return "generic";
493 switch(hostInfo.cpu_subtype) {
494 case CPU_SUBTYPE_POWERPC_601: return "601";
495 case CPU_SUBTYPE_POWERPC_602: return "602";
496 case CPU_SUBTYPE_POWERPC_603: return "603";
497 case CPU_SUBTYPE_POWERPC_603e: return "603e";
498 case CPU_SUBTYPE_POWERPC_603ev: return "603ev";
499 case CPU_SUBTYPE_POWERPC_604: return "604";
500 case CPU_SUBTYPE_POWERPC_604e: return "604e";
501 case CPU_SUBTYPE_POWERPC_620: return "620";
502 case CPU_SUBTYPE_POWERPC_750: return "750";
503 case CPU_SUBTYPE_POWERPC_7400: return "7400";
504 case CPU_SUBTYPE_POWERPC_7450: return "7450";
505 case CPU_SUBTYPE_POWERPC_970: return "970";
511 #elif defined(__linux__) && (defined(__ppc__) || defined(__powerpc__))
512 StringRef sys::getHostCPUName() {
513 // Access to the Processor Version Register (PVR) on PowerPC is privileged,
514 // and so we must use an operating-system interface to determine the current
515 // processor type. On Linux, this is exposed through the /proc/cpuinfo file.
516 const char *generic = "generic";
518 // The cpu line is second (after the 'processor: 0' line), so if this
519 // buffer is too small then something has changed (or is wrong).
521 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
522 if (CPUInfoSize == -1)
525 const char *CPUInfoStart = buffer;
526 const char *CPUInfoEnd = buffer + CPUInfoSize;
528 const char *CIP = CPUInfoStart;
530 const char *CPUStart = 0;
533 // We need to find the first line which starts with cpu, spaces, and a colon.
534 // After the colon, there may be some additional spaces and then the cpu type.
535 while (CIP < CPUInfoEnd && CPUStart == 0) {
536 if (CIP < CPUInfoEnd && *CIP == '\n')
539 if (CIP < CPUInfoEnd && *CIP == 'c') {
541 if (CIP < CPUInfoEnd && *CIP == 'p') {
543 if (CIP < CPUInfoEnd && *CIP == 'u') {
545 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
548 if (CIP < CPUInfoEnd && *CIP == ':') {
550 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
553 if (CIP < CPUInfoEnd) {
555 while (CIP < CPUInfoEnd && (*CIP != ' ' && *CIP != '\t' &&
556 *CIP != ',' && *CIP != '\n'))
558 CPULen = CIP - CPUStart;
566 while (CIP < CPUInfoEnd && *CIP != '\n')
573 return StringSwitch<const char *>(StringRef(CPUStart, CPULen))
574 .Case("604e", "604e")
576 .Case("7400", "7400")
577 .Case("7410", "7400")
578 .Case("7447", "7400")
579 .Case("7455", "7450")
581 .Case("POWER4", "970")
582 .Case("PPC970FX", "970")
583 .Case("PPC970MP", "970")
585 .Case("POWER5", "g5")
587 .Case("POWER6", "pwr6")
588 .Case("POWER7", "pwr7")
589 .Case("POWER8", "pwr8")
590 .Case("POWER8E", "pwr8")
593 #elif defined(__linux__) && defined(__arm__)
594 StringRef sys::getHostCPUName() {
595 // The cpuid register on arm is not accessible from user space. On Linux,
596 // it is exposed through the /proc/cpuinfo file.
598 // Read 1024 bytes from /proc/cpuinfo, which should contain the CPU part line
601 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
602 if (CPUInfoSize == -1)
605 StringRef Str(buffer, CPUInfoSize);
607 SmallVector<StringRef, 32> Lines;
608 Str.split(Lines, "\n");
610 // Look for the CPU implementer line.
611 StringRef Implementer;
612 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
613 if (Lines[I].startswith("CPU implementer"))
614 Implementer = Lines[I].substr(15).ltrim("\t :");
616 if (Implementer == "0x41") // ARM Ltd.
617 // Look for the CPU part line.
618 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
619 if (Lines[I].startswith("CPU part"))
620 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
621 // values correspond to the "Part number" in the CP15/c0 register. The
622 // contents are specified in the various processor manuals.
623 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :"))
624 .Case("0x926", "arm926ej-s")
625 .Case("0xb02", "mpcore")
626 .Case("0xb36", "arm1136j-s")
627 .Case("0xb56", "arm1156t2-s")
628 .Case("0xb76", "arm1176jz-s")
629 .Case("0xc08", "cortex-a8")
630 .Case("0xc09", "cortex-a9")
631 .Case("0xc0f", "cortex-a15")
632 .Case("0xc20", "cortex-m0")
633 .Case("0xc23", "cortex-m3")
634 .Case("0xc24", "cortex-m4")
637 if (Implementer == "0x51") // Qualcomm Technologies, Inc.
638 // Look for the CPU part line.
639 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
640 if (Lines[I].startswith("CPU part"))
641 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
642 // values correspond to the "Part number" in the CP15/c0 register. The
643 // contents are specified in the various processor manuals.
644 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :"))
645 .Case("0x06f", "krait") // APQ8064
650 #elif defined(__linux__) && defined(__s390x__)
651 StringRef sys::getHostCPUName() {
652 // STIDP is a privileged operation, so use /proc/cpuinfo instead.
654 // The "processor 0:" line comes after a fair amount of other information,
655 // including a cache breakdown, but this should be plenty.
657 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
658 if (CPUInfoSize == -1)
661 StringRef Str(buffer, CPUInfoSize);
662 SmallVector<StringRef, 32> Lines;
663 Str.split(Lines, "\n");
664 for (unsigned I = 0, E = Lines.size(); I != E; ++I) {
665 if (Lines[I].startswith("processor ")) {
666 size_t Pos = Lines[I].find("machine = ");
667 if (Pos != StringRef::npos) {
668 Pos += sizeof("machine = ") - 1;
670 if (!Lines[I].drop_front(Pos).getAsInteger(10, Id)) {
684 StringRef sys::getHostCPUName() {
689 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\
690 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
691 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
692 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
699 if (GetX86CpuIDAndInfo(0, &MaxLevel, text.u+0, text.u+2, text.u+1) ||
703 GetX86CpuIDAndInfo(1, &EAX, &EBX, &ECX, &EDX);
705 Features["cmov"] = (EDX >> 15) & 1;
706 Features["mmx"] = (EDX >> 23) & 1;
707 Features["sse"] = (EDX >> 25) & 1;
708 Features["sse2"] = (EDX >> 26) & 1;
709 Features["sse3"] = (ECX >> 0) & 1;
710 Features["ssse3"] = (ECX >> 9) & 1;
711 Features["sse4.1"] = (ECX >> 19) & 1;
712 Features["sse4.2"] = (ECX >> 20) & 1;
714 Features["pclmul"] = (ECX >> 1) & 1;
715 Features["fma"] = (ECX >> 12) & 1;
716 Features["cx16"] = (ECX >> 13) & 1;
717 Features["movbe"] = (ECX >> 22) & 1;
718 Features["popcnt"] = (ECX >> 23) & 1;
719 Features["aes"] = (ECX >> 25) & 1;
720 Features["f16c"] = (ECX >> 29) & 1;
721 Features["rdrnd"] = (ECX >> 30) & 1;
723 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
724 // indicates that the AVX registers will be saved and restored on context
725 // switch, then we have full AVX support.
726 bool HasAVX = ((ECX >> 27) & 1) && ((ECX >> 28) & 1) &&
727 !GetX86XCR0(&EAX, &EDX) && ((EAX & 0x6) == 0x6);
728 Features["avx"] = HasAVX;
730 // AVX512 requires additional context to be saved by the OS.
731 bool HasAVX512Save = HasAVX && ((EAX & 0xe0) == 0xe0);
733 unsigned MaxExtLevel;
734 GetX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX);
736 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 &&
737 !GetX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
738 Features["lzcnt"] = HasExtLeaf1 && ((ECX >> 5) & 1);
739 Features["sse4a"] = HasExtLeaf1 && ((ECX >> 6) & 1);
740 Features["prfchw"] = HasExtLeaf1 && ((ECX >> 8) & 1);
741 Features["xop"] = HasExtLeaf1 && ((ECX >> 11) & 1);
742 Features["fma4"] = HasExtLeaf1 && ((ECX >> 16) & 1);
743 Features["tbm"] = HasExtLeaf1 && ((ECX >> 21) & 1);
745 bool HasLeaf7 = MaxLevel >= 7 &&
746 !GetX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX);
748 // AVX2 is only supported if we have the OS save support from AVX.
749 Features["avx2"] = HasAVX && HasLeaf7 && (EBX >> 5) & 1;
751 Features["fsgsbase"] = HasLeaf7 && ((EBX >> 0) & 1);
752 Features["bmi"] = HasLeaf7 && ((EBX >> 3) & 1);
753 Features["hle"] = HasLeaf7 && ((EBX >> 4) & 1);
754 Features["bmi2"] = HasLeaf7 && ((EBX >> 8) & 1);
755 Features["rtm"] = HasLeaf7 && ((EBX >> 11) & 1);
756 Features["rdseed"] = HasLeaf7 && ((EBX >> 18) & 1);
757 Features["adx"] = HasLeaf7 && ((EBX >> 19) & 1);
758 Features["sha"] = HasLeaf7 && ((EBX >> 29) & 1);
760 // AVX512 is only supported if the OS supports the context save for it.
761 Features["avx512f"] = HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save;
762 Features["avx512dq"] = HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save;
763 Features["avx512pf"] = HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save;
764 Features["avx512er"] = HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save;
765 Features["avx512cd"] = HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save;
766 Features["avx512bw"] = HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save;
767 Features["avx512vl"] = HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save;
771 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
772 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
773 // Read 1024 bytes from /proc/cpuinfo, which should contain the Features line
776 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
777 if (CPUInfoSize == -1)
780 StringRef Str(buffer, CPUInfoSize);
782 SmallVector<StringRef, 32> Lines;
783 Str.split(Lines, "\n");
785 SmallVector<StringRef, 32> CPUFeatures;
787 // Look for the CPU features.
788 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
789 if (Lines[I].startswith("Features")) {
790 Lines[I].split(CPUFeatures, " ");
794 #if defined(__aarch64__)
795 // Keep track of which crypto features we have seen
805 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) {
806 StringRef LLVMFeatureStr = StringSwitch<StringRef>(CPUFeatures[I])
807 #if defined(__aarch64__)
808 .Case("asimd", "neon")
809 .Case("fp", "fp-armv8")
810 .Case("crc32", "crc")
812 .Case("half", "fp16")
813 .Case("neon", "neon")
814 .Case("vfpv3", "vfp3")
815 .Case("vfpv3d16", "d16")
816 .Case("vfpv4", "vfp4")
817 .Case("idiva", "hwdiv-arm")
818 .Case("idivt", "hwdiv")
822 #if defined(__aarch64__)
823 // We need to check crypto separately since we need all of the crypto
824 // extensions to enable the subtarget feature
825 if (CPUFeatures[I] == "aes")
827 else if (CPUFeatures[I] == "pmull")
829 else if (CPUFeatures[I] == "sha1")
831 else if (CPUFeatures[I] == "sha2")
835 if (LLVMFeatureStr != "")
836 Features[LLVMFeatureStr] = true;
839 #if defined(__aarch64__)
840 // If we have all crypto bits we can add the feature
841 if (crypto == (CAP_AES | CAP_PMULL | CAP_SHA1 | CAP_SHA2))
842 Features["crypto"] = true;
848 bool sys::getHostCPUFeatures(StringMap<bool> &Features){
853 std::string sys::getProcessTriple() {
854 Triple PT(Triple::normalize(LLVM_HOST_TRIPLE));
856 if (sizeof(void *) == 8 && PT.isArch32Bit())
857 PT = PT.get64BitArchVariant();
858 if (sizeof(void *) == 4 && PT.isArch64Bit())
859 PT = PT.get32BitArchVariant();