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);
229 GetX86CpuIDAndInfo(0, &MaxLeaf, text.u+0, text.u+2, text.u+1);
231 bool HasMMX = (EDX >> 23) & 1;
232 bool HasSSE = (EDX >> 25) & 1;
233 bool HasSSE2 = (EDX >> 26) & 1;
234 bool HasSSE3 = (ECX >> 0) & 1;
235 bool HasSSSE3 = (ECX >> 9) & 1;
236 bool HasSSE41 = (ECX >> 19) & 1;
237 bool HasSSE42 = (ECX >> 20) & 1;
238 bool HasMOVBE = (ECX >> 22) & 1;
239 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
240 // indicates that the AVX registers will be saved and restored on context
241 // switch, then we have full AVX support.
242 const unsigned AVXBits = (1 << 27) | (1 << 28);
243 bool HasAVX = ((ECX & AVXBits) == AVXBits) && !GetX86XCR0(&EAX, &EDX) &&
244 ((EAX & 0x6) == 0x6);
245 bool HasAVX512Save = HasAVX && ((EAX & 0xe0) == 0xe0);
246 bool HasLeaf7 = MaxLeaf >= 0x7 &&
247 !GetX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX);
248 bool HasADX = HasLeaf7 && ((EBX >> 19) & 1);
249 bool HasAVX2 = HasAVX && HasLeaf7 && (EBX & 0x20);
250 bool HasAVX512 = HasLeaf7 && HasAVX512Save && ((EBX >> 16) & 1);
252 GetX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
253 bool Em64T = (EDX >> 29) & 0x1;
254 bool HasTBM = (ECX >> 21) & 0x1;
256 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
262 case 0: // Intel486 DX processors
263 case 1: // Intel486 DX processors
264 case 2: // Intel486 SX processors
265 case 3: // Intel487 processors, IntelDX2 OverDrive processors,
266 // IntelDX2 processors
267 case 4: // Intel486 SL processor
268 case 5: // IntelSX2 processors
269 case 7: // Write-Back Enhanced IntelDX2 processors
270 case 8: // IntelDX4 OverDrive processors, IntelDX4 processors
271 default: return "i486";
275 case 1: // Pentium OverDrive processor for Pentium processor (60, 66),
276 // Pentium processors (60, 66)
277 case 2: // Pentium OverDrive processor for Pentium processor (75, 90,
278 // 100, 120, 133), Pentium processors (75, 90, 100, 120, 133,
280 case 3: // Pentium OverDrive processors for Intel486 processor-based
284 case 4: // Pentium OverDrive processor with MMX technology for Pentium
285 // processor (75, 90, 100, 120, 133), Pentium processor with
286 // MMX technology (166, 200)
287 return "pentium-mmx";
289 default: return "pentium";
293 case 1: // Pentium Pro processor
296 case 3: // Intel Pentium II OverDrive processor, Pentium II processor,
298 case 5: // Pentium II processor, model 05, Pentium II Xeon processor,
299 // model 05, and Intel Celeron processor, model 05
300 case 6: // Celeron processor, model 06
303 case 7: // Pentium III processor, model 07, and Pentium III Xeon
304 // processor, model 07
305 case 8: // Pentium III processor, model 08, Pentium III Xeon processor,
306 // model 08, and Celeron processor, model 08
307 case 10: // Pentium III Xeon processor, model 0Ah
308 case 11: // Pentium III processor, model 0Bh
311 case 9: // Intel Pentium M processor, Intel Celeron M processor model 09.
312 case 13: // Intel Pentium M processor, Intel Celeron M processor, model
313 // 0Dh. All processors are manufactured using the 90 nm process.
314 case 21: // Intel EP80579 Integrated Processor and Intel EP80579
315 // Integrated Processor with Intel QuickAssist Technology
318 case 14: // Intel Core Duo processor, Intel Core Solo processor, model
319 // 0Eh. All processors are manufactured using the 65 nm process.
322 case 15: // Intel Core 2 Duo processor, Intel Core 2 Duo mobile
323 // processor, Intel Core 2 Quad processor, Intel Core 2 Quad
324 // mobile processor, Intel Core 2 Extreme processor, Intel
325 // Pentium Dual-Core processor, Intel Xeon processor, model
326 // 0Fh. All processors are manufactured using the 65 nm process.
327 case 22: // Intel Celeron processor model 16h. All processors are
328 // manufactured using the 65 nm process
331 case 23: // Intel Core 2 Extreme processor, Intel Xeon processor, model
332 // 17h. All processors are manufactured using the 45 nm process.
334 // 45nm: Penryn , Wolfdale, Yorkfield (XE)
335 case 29: // Intel Xeon processor MP. All processors are manufactured using
336 // the 45 nm process.
339 case 26: // Intel Core i7 processor and Intel Xeon processor. All
340 // processors are manufactured using the 45 nm process.
341 case 30: // Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz.
342 // As found in a Summer 2010 model iMac.
343 case 46: // Nehalem EX
345 case 37: // Intel Core i7, laptop version.
346 case 44: // Intel Core i7 processor and Intel Xeon processor. All
347 // processors are manufactured using the 32 nm process.
348 case 47: // Westmere EX
352 case 42: // Intel Core i7 processor. All processors are manufactured
353 // using the 32 nm process.
355 return "sandybridge";
359 case 62: // Ivy Bridge EP
373 case 28: // Most 45 nm Intel Atom processors
374 case 38: // 45 nm Atom Lincroft
375 case 39: // 32 nm Atom Medfield
376 case 53: // 32 nm Atom Midview
377 case 54: // 32 nm Atom Midview
380 // Atom Silvermont codes from the Intel software optimization guide.
386 default: // Unknown family 6 CPU, try to guess.
394 return "sandybridge";
396 return HasMOVBE ? "silvermont" : "nehalem";
400 return HasMOVBE ? "bonnell" : "core2";
413 case 0: // Pentium 4 processor, Intel Xeon processor. All processors are
414 // model 00h and manufactured using the 0.18 micron process.
415 case 1: // Pentium 4 processor, Intel Xeon processor, Intel Xeon
416 // processor MP, and Intel Celeron processor. All processors are
417 // model 01h and manufactured using the 0.18 micron process.
418 case 2: // Pentium 4 processor, Mobile Intel Pentium 4 processor - M,
419 // Intel Xeon processor, Intel Xeon processor MP, Intel Celeron
420 // processor, and Mobile Intel Celeron processor. All processors
421 // are model 02h and manufactured using the 0.13 micron process.
422 return (Em64T) ? "x86-64" : "pentium4";
424 case 3: // Pentium 4 processor, Intel Xeon processor, Intel Celeron D
425 // processor. All processors are model 03h and manufactured using
426 // the 90 nm process.
427 case 4: // Pentium 4 processor, Pentium 4 processor Extreme Edition,
428 // Pentium D processor, Intel Xeon processor, Intel Xeon
429 // processor MP, Intel Celeron D processor. All processors are
430 // model 04h and manufactured using the 90 nm process.
431 case 6: // Pentium 4 processor, Pentium D processor, Pentium processor
432 // Extreme Edition, Intel Xeon processor, Intel Xeon processor
433 // MP, Intel Celeron D processor. All processors are model 06h
434 // and manufactured using the 65 nm process.
435 return (Em64T) ? "nocona" : "prescott";
438 return (Em64T) ? "x86-64" : "pentium4";
445 } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
446 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
447 // appears to be no way to generate the wide variety of AMD-specific targets
448 // from the information returned from CPUID.
456 case 8: return "k6-2";
458 case 13: return "k6-3";
459 case 10: return "geode";
460 default: return "pentium";
464 case 4: return "athlon-tbird";
467 case 8: return "athlon-mp";
468 case 10: return "athlon-xp";
469 default: return "athlon";
475 case 1: return "opteron";
476 case 5: return "athlon-fx"; // also opteron
477 default: return "athlon64";
484 if (!HasAVX) // If the OS doesn't support AVX provide a sane fallback.
487 return "bdver4"; // 50h-6Fh: Excavator
489 return "bdver3"; // 30h-3Fh: Steamroller
490 if (Model >= 0x10 || HasTBM)
491 return "bdver2"; // 10h-1Fh: Piledriver
492 return "bdver1"; // 00h-0Fh: Bulldozer
494 if (!HasAVX) // If the OS doesn't support AVX provide a sane fallback.
503 #elif defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
504 StringRef sys::getHostCPUName() {
505 host_basic_info_data_t hostInfo;
506 mach_msg_type_number_t infoCount;
508 infoCount = HOST_BASIC_INFO_COUNT;
509 host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&hostInfo,
512 if (hostInfo.cpu_type != CPU_TYPE_POWERPC) return "generic";
514 switch(hostInfo.cpu_subtype) {
515 case CPU_SUBTYPE_POWERPC_601: return "601";
516 case CPU_SUBTYPE_POWERPC_602: return "602";
517 case CPU_SUBTYPE_POWERPC_603: return "603";
518 case CPU_SUBTYPE_POWERPC_603e: return "603e";
519 case CPU_SUBTYPE_POWERPC_603ev: return "603ev";
520 case CPU_SUBTYPE_POWERPC_604: return "604";
521 case CPU_SUBTYPE_POWERPC_604e: return "604e";
522 case CPU_SUBTYPE_POWERPC_620: return "620";
523 case CPU_SUBTYPE_POWERPC_750: return "750";
524 case CPU_SUBTYPE_POWERPC_7400: return "7400";
525 case CPU_SUBTYPE_POWERPC_7450: return "7450";
526 case CPU_SUBTYPE_POWERPC_970: return "970";
532 #elif defined(__linux__) && (defined(__ppc__) || defined(__powerpc__))
533 StringRef sys::getHostCPUName() {
534 // Access to the Processor Version Register (PVR) on PowerPC is privileged,
535 // and so we must use an operating-system interface to determine the current
536 // processor type. On Linux, this is exposed through the /proc/cpuinfo file.
537 const char *generic = "generic";
539 // The cpu line is second (after the 'processor: 0' line), so if this
540 // buffer is too small then something has changed (or is wrong).
542 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
543 if (CPUInfoSize == -1)
546 const char *CPUInfoStart = buffer;
547 const char *CPUInfoEnd = buffer + CPUInfoSize;
549 const char *CIP = CPUInfoStart;
551 const char *CPUStart = 0;
554 // We need to find the first line which starts with cpu, spaces, and a colon.
555 // After the colon, there may be some additional spaces and then the cpu type.
556 while (CIP < CPUInfoEnd && CPUStart == 0) {
557 if (CIP < CPUInfoEnd && *CIP == '\n')
560 if (CIP < CPUInfoEnd && *CIP == 'c') {
562 if (CIP < CPUInfoEnd && *CIP == 'p') {
564 if (CIP < CPUInfoEnd && *CIP == 'u') {
566 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
569 if (CIP < CPUInfoEnd && *CIP == ':') {
571 while (CIP < CPUInfoEnd && (*CIP == ' ' || *CIP == '\t'))
574 if (CIP < CPUInfoEnd) {
576 while (CIP < CPUInfoEnd && (*CIP != ' ' && *CIP != '\t' &&
577 *CIP != ',' && *CIP != '\n'))
579 CPULen = CIP - CPUStart;
587 while (CIP < CPUInfoEnd && *CIP != '\n')
594 return StringSwitch<const char *>(StringRef(CPUStart, CPULen))
595 .Case("604e", "604e")
597 .Case("7400", "7400")
598 .Case("7410", "7400")
599 .Case("7447", "7400")
600 .Case("7455", "7450")
602 .Case("POWER4", "970")
603 .Case("PPC970FX", "970")
604 .Case("PPC970MP", "970")
606 .Case("POWER5", "g5")
608 .Case("POWER6", "pwr6")
609 .Case("POWER7", "pwr7")
610 .Case("POWER8", "pwr8")
611 .Case("POWER8E", "pwr8")
614 #elif defined(__linux__) && defined(__arm__)
615 StringRef sys::getHostCPUName() {
616 // The cpuid register on arm is not accessible from user space. On Linux,
617 // it is exposed through the /proc/cpuinfo file.
619 // Read 1024 bytes from /proc/cpuinfo, which should contain the CPU part line
622 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
623 if (CPUInfoSize == -1)
626 StringRef Str(buffer, CPUInfoSize);
628 SmallVector<StringRef, 32> Lines;
629 Str.split(Lines, "\n");
631 // Look for the CPU implementer line.
632 StringRef Implementer;
633 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
634 if (Lines[I].startswith("CPU implementer"))
635 Implementer = Lines[I].substr(15).ltrim("\t :");
637 if (Implementer == "0x41") // ARM Ltd.
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("0x926", "arm926ej-s")
646 .Case("0xb02", "mpcore")
647 .Case("0xb36", "arm1136j-s")
648 .Case("0xb56", "arm1156t2-s")
649 .Case("0xb76", "arm1176jz-s")
650 .Case("0xc08", "cortex-a8")
651 .Case("0xc09", "cortex-a9")
652 .Case("0xc0f", "cortex-a15")
653 .Case("0xc20", "cortex-m0")
654 .Case("0xc23", "cortex-m3")
655 .Case("0xc24", "cortex-m4")
658 if (Implementer == "0x51") // Qualcomm Technologies, Inc.
659 // Look for the CPU part line.
660 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
661 if (Lines[I].startswith("CPU part"))
662 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
663 // values correspond to the "Part number" in the CP15/c0 register. The
664 // contents are specified in the various processor manuals.
665 return StringSwitch<const char *>(Lines[I].substr(8).ltrim("\t :"))
666 .Case("0x06f", "krait") // APQ8064
671 #elif defined(__linux__) && defined(__s390x__)
672 StringRef sys::getHostCPUName() {
673 // STIDP is a privileged operation, so use /proc/cpuinfo instead.
675 // The "processor 0:" line comes after a fair amount of other information,
676 // including a cache breakdown, but this should be plenty.
678 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
679 if (CPUInfoSize == -1)
682 StringRef Str(buffer, CPUInfoSize);
683 SmallVector<StringRef, 32> Lines;
684 Str.split(Lines, "\n");
686 // Look for the CPU features.
687 SmallVector<StringRef, 32> CPUFeatures;
688 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
689 if (Lines[I].startswith("features")) {
690 size_t Pos = Lines[I].find(":");
691 if (Pos != StringRef::npos) {
692 Lines[I].drop_front(Pos + 1).split(CPUFeatures, " ");
697 // We need to check for the presence of vector support independently of
698 // the machine type, since we may only use the vector register set when
699 // supported by the kernel (and hypervisor).
700 bool HaveVectorSupport = false;
701 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) {
702 if (CPUFeatures[I] == "vx")
703 HaveVectorSupport = true;
706 // Now check the processor machine type.
707 for (unsigned I = 0, E = Lines.size(); I != E; ++I) {
708 if (Lines[I].startswith("processor ")) {
709 size_t Pos = Lines[I].find("machine = ");
710 if (Pos != StringRef::npos) {
711 Pos += sizeof("machine = ") - 1;
713 if (!Lines[I].drop_front(Pos).getAsInteger(10, Id)) {
714 if (Id >= 2964 && HaveVectorSupport)
729 StringRef sys::getHostCPUName() {
734 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\
735 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
736 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
737 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
744 if (GetX86CpuIDAndInfo(0, &MaxLevel, text.u+0, text.u+2, text.u+1) ||
748 GetX86CpuIDAndInfo(1, &EAX, &EBX, &ECX, &EDX);
750 Features["cmov"] = (EDX >> 15) & 1;
751 Features["mmx"] = (EDX >> 23) & 1;
752 Features["sse"] = (EDX >> 25) & 1;
753 Features["sse2"] = (EDX >> 26) & 1;
754 Features["sse3"] = (ECX >> 0) & 1;
755 Features["ssse3"] = (ECX >> 9) & 1;
756 Features["sse4.1"] = (ECX >> 19) & 1;
757 Features["sse4.2"] = (ECX >> 20) & 1;
759 Features["pclmul"] = (ECX >> 1) & 1;
760 Features["cx16"] = (ECX >> 13) & 1;
761 Features["movbe"] = (ECX >> 22) & 1;
762 Features["popcnt"] = (ECX >> 23) & 1;
763 Features["aes"] = (ECX >> 25) & 1;
764 Features["rdrnd"] = (ECX >> 30) & 1;
766 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
767 // indicates that the AVX registers will be saved and restored on context
768 // switch, then we have full AVX support.
769 bool HasAVX = ((ECX >> 27) & 1) && ((ECX >> 28) & 1) &&
770 !GetX86XCR0(&EAX, &EDX) && ((EAX & 0x6) == 0x6);
771 Features["avx"] = HasAVX;
772 Features["fma"] = HasAVX && (ECX >> 12) & 1;
773 Features["f16c"] = HasAVX && (ECX >> 29) & 1;
775 // AVX512 requires additional context to be saved by the OS.
776 bool HasAVX512Save = HasAVX && ((EAX & 0xe0) == 0xe0);
778 unsigned MaxExtLevel;
779 GetX86CpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX);
781 bool HasExtLeaf1 = MaxExtLevel >= 0x80000001 &&
782 !GetX86CpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
783 Features["lzcnt"] = HasExtLeaf1 && ((ECX >> 5) & 1);
784 Features["sse4a"] = HasExtLeaf1 && ((ECX >> 6) & 1);
785 Features["prfchw"] = HasExtLeaf1 && ((ECX >> 8) & 1);
786 Features["xop"] = HasAVX && HasExtLeaf1 && ((ECX >> 11) & 1);
787 Features["fma4"] = HasAVX && HasExtLeaf1 && ((ECX >> 16) & 1);
788 Features["tbm"] = HasExtLeaf1 && ((ECX >> 21) & 1);
790 bool HasLeaf7 = MaxLevel >= 7 &&
791 !GetX86CpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX);
793 // AVX2 is only supported if we have the OS save support from AVX.
794 Features["avx2"] = HasAVX && HasLeaf7 && (EBX >> 5) & 1;
796 Features["fsgsbase"] = HasLeaf7 && ((EBX >> 0) & 1);
797 Features["bmi"] = HasLeaf7 && ((EBX >> 3) & 1);
798 Features["hle"] = HasLeaf7 && ((EBX >> 4) & 1);
799 Features["bmi2"] = HasLeaf7 && ((EBX >> 8) & 1);
800 Features["rtm"] = HasLeaf7 && ((EBX >> 11) & 1);
801 Features["rdseed"] = HasLeaf7 && ((EBX >> 18) & 1);
802 Features["adx"] = HasLeaf7 && ((EBX >> 19) & 1);
803 Features["sha"] = HasLeaf7 && ((EBX >> 29) & 1);
805 // AVX512 is only supported if the OS supports the context save for it.
806 Features["avx512f"] = HasLeaf7 && ((EBX >> 16) & 1) && HasAVX512Save;
807 Features["avx512dq"] = HasLeaf7 && ((EBX >> 17) & 1) && HasAVX512Save;
808 Features["avx512pf"] = HasLeaf7 && ((EBX >> 26) & 1) && HasAVX512Save;
809 Features["avx512er"] = HasLeaf7 && ((EBX >> 27) & 1) && HasAVX512Save;
810 Features["avx512cd"] = HasLeaf7 && ((EBX >> 28) & 1) && HasAVX512Save;
811 Features["avx512bw"] = HasLeaf7 && ((EBX >> 30) & 1) && HasAVX512Save;
812 Features["avx512vl"] = HasLeaf7 && ((EBX >> 31) & 1) && HasAVX512Save;
816 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
817 bool sys::getHostCPUFeatures(StringMap<bool> &Features) {
818 // Read 1024 bytes from /proc/cpuinfo, which should contain the Features line
821 ssize_t CPUInfoSize = readCpuInfo(buffer, sizeof(buffer));
822 if (CPUInfoSize == -1)
825 StringRef Str(buffer, CPUInfoSize);
827 SmallVector<StringRef, 32> Lines;
828 Str.split(Lines, "\n");
830 SmallVector<StringRef, 32> CPUFeatures;
832 // Look for the CPU features.
833 for (unsigned I = 0, E = Lines.size(); I != E; ++I)
834 if (Lines[I].startswith("Features")) {
835 Lines[I].split(CPUFeatures, " ");
839 #if defined(__aarch64__)
840 // Keep track of which crypto features we have seen
850 for (unsigned I = 0, E = CPUFeatures.size(); I != E; ++I) {
851 StringRef LLVMFeatureStr = StringSwitch<StringRef>(CPUFeatures[I])
852 #if defined(__aarch64__)
853 .Case("asimd", "neon")
854 .Case("fp", "fp-armv8")
855 .Case("crc32", "crc")
857 .Case("half", "fp16")
858 .Case("neon", "neon")
859 .Case("vfpv3", "vfp3")
860 .Case("vfpv3d16", "d16")
861 .Case("vfpv4", "vfp4")
862 .Case("idiva", "hwdiv-arm")
863 .Case("idivt", "hwdiv")
867 #if defined(__aarch64__)
868 // We need to check crypto separately since we need all of the crypto
869 // extensions to enable the subtarget feature
870 if (CPUFeatures[I] == "aes")
872 else if (CPUFeatures[I] == "pmull")
874 else if (CPUFeatures[I] == "sha1")
876 else if (CPUFeatures[I] == "sha2")
880 if (LLVMFeatureStr != "")
881 Features[LLVMFeatureStr] = true;
884 #if defined(__aarch64__)
885 // If we have all crypto bits we can add the feature
886 if (crypto == (CAP_AES | CAP_PMULL | CAP_SHA1 | CAP_SHA2))
887 Features["crypto"] = true;
893 bool sys::getHostCPUFeatures(StringMap<bool> &Features){
898 std::string sys::getProcessTriple() {
899 Triple PT(Triple::normalize(LLVM_HOST_TRIPLE));
901 if (sizeof(void *) == 8 && PT.isArch32Bit())
902 PT = PT.get64BitArchVariant();
903 if (sizeof(void *) == 4 && PT.isArch64Bit())
904 PT = PT.get32BitArchVariant();