1 //===-- Host.cpp - Implement OS Host Concept --------------------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements the operating system Host concept.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/Support/Host.h"
14 #include "llvm/Support/TargetParser.h"
15 #include "llvm/ADT/SmallSet.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/StringRef.h"
18 #include "llvm/ADT/StringSwitch.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/Config/llvm-config.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/FileSystem.h"
23 #include "llvm/Support/MemoryBuffer.h"
24 #include "llvm/Support/raw_ostream.h"
28 // Include the platform-specific parts of this class.
30 #include "Unix/Host.inc"
33 #include "Windows/Host.inc"
38 #if defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
39 #include <mach/host_info.h>
40 #include <mach/mach.h>
41 #include <mach/mach_host.h>
42 #include <mach/machine.h>
45 #define DEBUG_TYPE "host-detection"
47 //===----------------------------------------------------------------------===//
49 // Implementations of the CPU detection routines
51 //===----------------------------------------------------------------------===//
55 static std::unique_ptr
<llvm::MemoryBuffer
>
56 LLVM_ATTRIBUTE_UNUSED
getProcCpuinfoContent() {
57 llvm::ErrorOr
<std::unique_ptr
<llvm::MemoryBuffer
>> Text
=
58 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo");
59 if (std::error_code EC
= Text
.getError()) {
60 llvm::errs() << "Can't read "
61 << "/proc/cpuinfo: " << EC
.message() << "\n";
64 return std::move(*Text
);
67 StringRef
sys::detail::getHostCPUNameForPowerPC(StringRef ProcCpuinfoContent
) {
68 // Access to the Processor Version Register (PVR) on PowerPC is privileged,
69 // and so we must use an operating-system interface to determine the current
70 // processor type. On Linux, this is exposed through the /proc/cpuinfo file.
71 const char *generic
= "generic";
73 // The cpu line is second (after the 'processor: 0' line), so if this
74 // buffer is too small then something has changed (or is wrong).
75 StringRef::const_iterator CPUInfoStart
= ProcCpuinfoContent
.begin();
76 StringRef::const_iterator CPUInfoEnd
= ProcCpuinfoContent
.end();
78 StringRef::const_iterator CIP
= CPUInfoStart
;
80 StringRef::const_iterator CPUStart
= 0;
83 // We need to find the first line which starts with cpu, spaces, and a colon.
84 // After the colon, there may be some additional spaces and then the cpu type.
85 while (CIP
< CPUInfoEnd
&& CPUStart
== 0) {
86 if (CIP
< CPUInfoEnd
&& *CIP
== '\n')
89 if (CIP
< CPUInfoEnd
&& *CIP
== 'c') {
91 if (CIP
< CPUInfoEnd
&& *CIP
== 'p') {
93 if (CIP
< CPUInfoEnd
&& *CIP
== 'u') {
95 while (CIP
< CPUInfoEnd
&& (*CIP
== ' ' || *CIP
== '\t'))
98 if (CIP
< CPUInfoEnd
&& *CIP
== ':') {
100 while (CIP
< CPUInfoEnd
&& (*CIP
== ' ' || *CIP
== '\t'))
103 if (CIP
< CPUInfoEnd
) {
105 while (CIP
< CPUInfoEnd
&& (*CIP
!= ' ' && *CIP
!= '\t' &&
106 *CIP
!= ',' && *CIP
!= '\n'))
108 CPULen
= CIP
- CPUStart
;
116 while (CIP
< CPUInfoEnd
&& *CIP
!= '\n')
123 return StringSwitch
<const char *>(StringRef(CPUStart
, CPULen
))
124 .Case("604e", "604e")
126 .Case("7400", "7400")
127 .Case("7410", "7400")
128 .Case("7447", "7400")
129 .Case("7455", "7450")
131 .Case("POWER4", "970")
132 .Case("PPC970FX", "970")
133 .Case("PPC970MP", "970")
135 .Case("POWER5", "g5")
137 .Case("POWER6", "pwr6")
138 .Case("POWER7", "pwr7")
139 .Case("POWER8", "pwr8")
140 .Case("POWER8E", "pwr8")
141 .Case("POWER8NVL", "pwr8")
142 .Case("POWER9", "pwr9")
146 StringRef
sys::detail::getHostCPUNameForARM(StringRef ProcCpuinfoContent
) {
147 // The cpuid register on arm is not accessible from user space. On Linux,
148 // it is exposed through the /proc/cpuinfo file.
150 // Read 32 lines from /proc/cpuinfo, which should contain the CPU part line
152 SmallVector
<StringRef
, 32> Lines
;
153 ProcCpuinfoContent
.split(Lines
, "\n");
155 // Look for the CPU implementer line.
156 StringRef Implementer
;
158 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
) {
159 if (Lines
[I
].startswith("CPU implementer"))
160 Implementer
= Lines
[I
].substr(15).ltrim("\t :");
161 if (Lines
[I
].startswith("Hardware"))
162 Hardware
= Lines
[I
].substr(8).ltrim("\t :");
165 if (Implementer
== "0x41") { // ARM Ltd.
166 // MSM8992/8994 may give cpu part for the core that the kernel is running on,
167 // which is undeterministic and wrong. Always return cortex-a53 for these SoC.
168 if (Hardware
.endswith("MSM8994") || Hardware
.endswith("MSM8996"))
172 // Look for the CPU part line.
173 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
)
174 if (Lines
[I
].startswith("CPU part"))
175 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
176 // values correspond to the "Part number" in the CP15/c0 register. The
177 // contents are specified in the various processor manuals.
178 return StringSwitch
<const char *>(Lines
[I
].substr(8).ltrim("\t :"))
179 .Case("0x926", "arm926ej-s")
180 .Case("0xb02", "mpcore")
181 .Case("0xb36", "arm1136j-s")
182 .Case("0xb56", "arm1156t2-s")
183 .Case("0xb76", "arm1176jz-s")
184 .Case("0xc08", "cortex-a8")
185 .Case("0xc09", "cortex-a9")
186 .Case("0xc0f", "cortex-a15")
187 .Case("0xc20", "cortex-m0")
188 .Case("0xc23", "cortex-m3")
189 .Case("0xc24", "cortex-m4")
190 .Case("0xd04", "cortex-a35")
191 .Case("0xd03", "cortex-a53")
192 .Case("0xd07", "cortex-a57")
193 .Case("0xd08", "cortex-a72")
194 .Case("0xd09", "cortex-a73")
195 .Case("0xd0a", "cortex-a75")
196 .Case("0xd0b", "cortex-a76")
200 if (Implementer
== "0x42" || Implementer
== "0x43") { // Broadcom | Cavium.
201 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
) {
202 if (Lines
[I
].startswith("CPU part")) {
203 return StringSwitch
<const char *>(Lines
[I
].substr(8).ltrim("\t :"))
204 .Case("0x516", "thunderx2t99")
205 .Case("0x0516", "thunderx2t99")
206 .Case("0xaf", "thunderx2t99")
207 .Case("0x0af", "thunderx2t99")
208 .Case("0xa1", "thunderxt88")
209 .Case("0x0a1", "thunderxt88")
215 if (Implementer
== "0x48") // HiSilicon Technologies, Inc.
216 // Look for the CPU part line.
217 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
)
218 if (Lines
[I
].startswith("CPU part"))
219 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
220 // values correspond to the "Part number" in the CP15/c0 register. The
221 // contents are specified in the various processor manuals.
222 return StringSwitch
<const char *>(Lines
[I
].substr(8).ltrim("\t :"))
223 .Case("0xd01", "tsv110")
226 if (Implementer
== "0x51") // Qualcomm Technologies, Inc.
227 // Look for the CPU part line.
228 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
)
229 if (Lines
[I
].startswith("CPU part"))
230 // The CPU part is a 3 digit hexadecimal number with a 0x prefix. The
231 // values correspond to the "Part number" in the CP15/c0 register. The
232 // contents are specified in the various processor manuals.
233 return StringSwitch
<const char *>(Lines
[I
].substr(8).ltrim("\t :"))
234 .Case("0x06f", "krait") // APQ8064
235 .Case("0x201", "kryo")
236 .Case("0x205", "kryo")
237 .Case("0x211", "kryo")
238 .Case("0x800", "cortex-a73")
239 .Case("0x801", "cortex-a73")
240 .Case("0x802", "cortex-a73")
241 .Case("0x803", "cortex-a73")
242 .Case("0x804", "cortex-a73")
243 .Case("0x805", "cortex-a73")
244 .Case("0xc00", "falkor")
245 .Case("0xc01", "saphira")
248 if (Implementer
== "0x53") { // Samsung Electronics Co., Ltd.
249 // The Exynos chips have a convoluted ID scheme that doesn't seem to follow
250 // any predictive pattern across variants and parts.
251 unsigned Variant
= 0, Part
= 0;
253 // Look for the CPU variant line, whose value is a 1 digit hexadecimal
254 // number, corresponding to the Variant bits in the CP15/C0 register.
256 if (I
.consume_front("CPU variant"))
257 I
.ltrim("\t :").getAsInteger(0, Variant
);
259 // Look for the CPU part line, whose value is a 3 digit hexadecimal
260 // number, corresponding to the PartNum bits in the CP15/C0 register.
262 if (I
.consume_front("CPU part"))
263 I
.ltrim("\t :").getAsInteger(0, Part
);
265 unsigned Exynos
= (Variant
<< 12) | Part
;
268 // Default by falling through to Exynos M1.
282 StringRef
sys::detail::getHostCPUNameForS390x(StringRef ProcCpuinfoContent
) {
283 // STIDP is a privileged operation, so use /proc/cpuinfo instead.
285 // The "processor 0:" line comes after a fair amount of other information,
286 // including a cache breakdown, but this should be plenty.
287 SmallVector
<StringRef
, 32> Lines
;
288 ProcCpuinfoContent
.split(Lines
, "\n");
290 // Look for the CPU features.
291 SmallVector
<StringRef
, 32> CPUFeatures
;
292 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
)
293 if (Lines
[I
].startswith("features")) {
294 size_t Pos
= Lines
[I
].find(":");
295 if (Pos
!= StringRef::npos
) {
296 Lines
[I
].drop_front(Pos
+ 1).split(CPUFeatures
, ' ');
301 // We need to check for the presence of vector support independently of
302 // the machine type, since we may only use the vector register set when
303 // supported by the kernel (and hypervisor).
304 bool HaveVectorSupport
= false;
305 for (unsigned I
= 0, E
= CPUFeatures
.size(); I
!= E
; ++I
) {
306 if (CPUFeatures
[I
] == "vx")
307 HaveVectorSupport
= true;
310 // Now check the processor machine type.
311 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
) {
312 if (Lines
[I
].startswith("processor ")) {
313 size_t Pos
= Lines
[I
].find("machine = ");
314 if (Pos
!= StringRef::npos
) {
315 Pos
+= sizeof("machine = ") - 1;
317 if (!Lines
[I
].drop_front(Pos
).getAsInteger(10, Id
)) {
318 if (Id
>= 8561 && HaveVectorSupport
)
320 if (Id
>= 3906 && HaveVectorSupport
)
322 if (Id
>= 2964 && HaveVectorSupport
)
337 StringRef
sys::detail::getHostCPUNameForBPF() {
338 #if !defined(__linux__) || !defined(__x86_64__)
341 uint8_t v3_insns
[40] __attribute__ ((aligned (8))) =
342 /* BPF_MOV64_IMM(BPF_REG_0, 0) */
343 { 0xb7, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
344 /* BPF_MOV64_IMM(BPF_REG_2, 1) */
345 0xb7, 0x2, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
346 /* BPF_JMP32_REG(BPF_JLT, BPF_REG_0, BPF_REG_2, 1) */
347 0xae, 0x20, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0,
348 /* BPF_MOV64_IMM(BPF_REG_0, 1) */
349 0xb7, 0x0, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
350 /* BPF_EXIT_INSN() */
351 0x95, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 };
353 uint8_t v2_insns
[40] __attribute__ ((aligned (8))) =
354 /* BPF_MOV64_IMM(BPF_REG_0, 0) */
355 { 0xb7, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
356 /* BPF_MOV64_IMM(BPF_REG_2, 1) */
357 0xb7, 0x2, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
358 /* BPF_JMP_REG(BPF_JLT, BPF_REG_0, BPF_REG_2, 1) */
359 0xad, 0x20, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0,
360 /* BPF_MOV64_IMM(BPF_REG_0, 1) */
361 0xb7, 0x0, 0x0, 0x0, 0x1, 0x0, 0x0, 0x0,
362 /* BPF_EXIT_INSN() */
363 0x95, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 };
365 struct bpf_prog_load_attr
{
373 uint32_t kern_version
;
376 attr
.prog_type
= 1; /* BPF_PROG_TYPE_SOCKET_FILTER */
378 attr
.insns
= (uint64_t)v3_insns
;
379 attr
.license
= (uint64_t)"DUMMY";
381 int fd
= syscall(321 /* __NR_bpf */, 5 /* BPF_PROG_LOAD */, &attr
,
388 /* Clear the whole attr in case its content changed by syscall. */
389 memset(&attr
, 0, sizeof(attr
));
390 attr
.prog_type
= 1; /* BPF_PROG_TYPE_SOCKET_FILTER */
392 attr
.insns
= (uint64_t)v2_insns
;
393 attr
.license
= (uint64_t)"DUMMY";
394 fd
= syscall(321 /* __NR_bpf */, 5 /* BPF_PROG_LOAD */, &attr
, sizeof(attr
));
403 #if defined(__i386__) || defined(_M_IX86) || \
404 defined(__x86_64__) || defined(_M_X64)
406 enum VendorSignatures
{
407 SIG_INTEL
= 0x756e6547 /* Genu */,
408 SIG_AMD
= 0x68747541 /* Auth */
411 // The check below for i386 was copied from clang's cpuid.h (__get_cpuid_max).
412 // Check motivated by bug reports for OpenSSL crashing on CPUs without CPUID
413 // support. Consequently, for i386, the presence of CPUID is checked first
414 // via the corresponding eflags bit.
415 // Removal of cpuid.h header motivated by PR30384
416 // Header cpuid.h and method __get_cpuid_max are not used in llvm, clang, openmp
417 // or test-suite, but are used in external projects e.g. libstdcxx
418 static bool isCpuIdSupported() {
419 #if defined(__GNUC__) || defined(__clang__)
420 #if defined(__i386__)
421 int __cpuid_supported
;
424 " movl %%eax,%%ecx\n"
425 " xorl $0x00200000,%%eax\n"
431 " cmpl %%eax,%%ecx\n"
435 : "=r"(__cpuid_supported
)
438 if (!__cpuid_supported
)
446 /// getX86CpuIDAndInfo - Execute the specified cpuid and return the 4 values in
447 /// the specified arguments. If we can't run cpuid on the host, return true.
448 static bool getX86CpuIDAndInfo(unsigned value
, unsigned *rEAX
, unsigned *rEBX
,
449 unsigned *rECX
, unsigned *rEDX
) {
450 #if defined(__GNUC__) || defined(__clang__)
451 #if defined(__x86_64__)
452 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually.
453 // FIXME: should we save this for Clang?
454 __asm__("movq\t%%rbx, %%rsi\n\t"
456 "xchgq\t%%rbx, %%rsi\n\t"
457 : "=a"(*rEAX
), "=S"(*rEBX
), "=c"(*rECX
), "=d"(*rEDX
)
460 #elif defined(__i386__)
461 __asm__("movl\t%%ebx, %%esi\n\t"
463 "xchgl\t%%ebx, %%esi\n\t"
464 : "=a"(*rEAX
), "=S"(*rEBX
), "=c"(*rECX
), "=d"(*rEDX
)
470 #elif defined(_MSC_VER)
471 // The MSVC intrinsic is portable across x86 and x64.
473 __cpuid(registers
, value
);
474 *rEAX
= registers
[0];
475 *rEBX
= registers
[1];
476 *rECX
= registers
[2];
477 *rEDX
= registers
[3];
484 /// getX86CpuIDAndInfoEx - Execute the specified cpuid with subleaf and return
485 /// the 4 values in the specified arguments. If we can't run cpuid on the host,
487 static bool getX86CpuIDAndInfoEx(unsigned value
, unsigned subleaf
,
488 unsigned *rEAX
, unsigned *rEBX
, unsigned *rECX
,
490 #if defined(__GNUC__) || defined(__clang__)
491 #if defined(__x86_64__)
492 // gcc doesn't know cpuid would clobber ebx/rbx. Preserve it manually.
493 // FIXME: should we save this for Clang?
494 __asm__("movq\t%%rbx, %%rsi\n\t"
496 "xchgq\t%%rbx, %%rsi\n\t"
497 : "=a"(*rEAX
), "=S"(*rEBX
), "=c"(*rECX
), "=d"(*rEDX
)
498 : "a"(value
), "c"(subleaf
));
500 #elif defined(__i386__)
501 __asm__("movl\t%%ebx, %%esi\n\t"
503 "xchgl\t%%ebx, %%esi\n\t"
504 : "=a"(*rEAX
), "=S"(*rEBX
), "=c"(*rECX
), "=d"(*rEDX
)
505 : "a"(value
), "c"(subleaf
));
510 #elif defined(_MSC_VER)
512 __cpuidex(registers
, value
, subleaf
);
513 *rEAX
= registers
[0];
514 *rEBX
= registers
[1];
515 *rECX
= registers
[2];
516 *rEDX
= registers
[3];
523 // Read control register 0 (XCR0). Used to detect features such as AVX.
524 static bool getX86XCR0(unsigned *rEAX
, unsigned *rEDX
) {
525 #if defined(__GNUC__) || defined(__clang__)
526 // Check xgetbv; this uses a .byte sequence instead of the instruction
527 // directly because older assemblers do not include support for xgetbv and
528 // there is no easy way to conditionally compile based on the assembler used.
529 __asm__(".byte 0x0f, 0x01, 0xd0" : "=a"(*rEAX
), "=d"(*rEDX
) : "c"(0));
531 #elif defined(_MSC_FULL_VER) && defined(_XCR_XFEATURE_ENABLED_MASK)
532 unsigned long long Result
= _xgetbv(_XCR_XFEATURE_ENABLED_MASK
);
534 *rEDX
= Result
>> 32;
541 static void detectX86FamilyModel(unsigned EAX
, unsigned *Family
,
543 *Family
= (EAX
>> 8) & 0xf; // Bits 8 - 11
544 *Model
= (EAX
>> 4) & 0xf; // Bits 4 - 7
545 if (*Family
== 6 || *Family
== 0xf) {
547 // Examine extended family ID if family ID is F.
548 *Family
+= (EAX
>> 20) & 0xff; // Bits 20 - 27
549 // Examine extended model ID if family ID is 6 or F.
550 *Model
+= ((EAX
>> 16) & 0xf) << 4; // Bits 16 - 19
555 getIntelProcessorTypeAndSubtype(unsigned Family
, unsigned Model
,
556 unsigned Brand_id
, unsigned Features
,
557 unsigned Features2
, unsigned Features3
,
558 unsigned *Type
, unsigned *Subtype
) {
563 *Type
= X86::INTEL_i386
;
566 *Type
= X86::INTEL_i486
;
569 if (Features
& (1 << X86::FEATURE_MMX
)) {
570 *Type
= X86::INTEL_PENTIUM_MMX
;
573 *Type
= X86::INTEL_PENTIUM
;
577 case 0x01: // Pentium Pro processor
578 *Type
= X86::INTEL_PENTIUM_PRO
;
580 case 0x03: // Intel Pentium II OverDrive processor, Pentium II processor,
582 case 0x05: // Pentium II processor, model 05, Pentium II Xeon processor,
583 // model 05, and Intel Celeron processor, model 05
584 case 0x06: // Celeron processor, model 06
585 *Type
= X86::INTEL_PENTIUM_II
;
587 case 0x07: // Pentium III processor, model 07, and Pentium III Xeon
588 // processor, model 07
589 case 0x08: // Pentium III processor, model 08, Pentium III Xeon processor,
590 // model 08, and Celeron processor, model 08
591 case 0x0a: // Pentium III Xeon processor, model 0Ah
592 case 0x0b: // Pentium III processor, model 0Bh
593 *Type
= X86::INTEL_PENTIUM_III
;
595 case 0x09: // Intel Pentium M processor, Intel Celeron M processor model 09.
596 case 0x0d: // Intel Pentium M processor, Intel Celeron M processor, model
597 // 0Dh. All processors are manufactured using the 90 nm process.
598 case 0x15: // Intel EP80579 Integrated Processor and Intel EP80579
599 // Integrated Processor with Intel QuickAssist Technology
600 *Type
= X86::INTEL_PENTIUM_M
;
602 case 0x0e: // Intel Core Duo processor, Intel Core Solo processor, model
603 // 0Eh. All processors are manufactured using the 65 nm process.
604 *Type
= X86::INTEL_CORE_DUO
;
606 case 0x0f: // Intel Core 2 Duo processor, Intel Core 2 Duo mobile
607 // processor, Intel Core 2 Quad processor, Intel Core 2 Quad
608 // mobile processor, Intel Core 2 Extreme processor, Intel
609 // Pentium Dual-Core processor, Intel Xeon processor, model
610 // 0Fh. All processors are manufactured using the 65 nm process.
611 case 0x16: // Intel Celeron processor model 16h. All processors are
612 // manufactured using the 65 nm process
613 *Type
= X86::INTEL_CORE2
; // "core2"
614 *Subtype
= X86::INTEL_CORE2_65
;
616 case 0x17: // Intel Core 2 Extreme processor, Intel Xeon processor, model
617 // 17h. All processors are manufactured using the 45 nm process.
619 // 45nm: Penryn , Wolfdale, Yorkfield (XE)
620 case 0x1d: // Intel Xeon processor MP. All processors are manufactured using
621 // the 45 nm process.
622 *Type
= X86::INTEL_CORE2
; // "penryn"
623 *Subtype
= X86::INTEL_CORE2_45
;
625 case 0x1a: // Intel Core i7 processor and Intel Xeon processor. All
626 // processors are manufactured using the 45 nm process.
627 case 0x1e: // Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz.
628 // As found in a Summer 2010 model iMac.
630 case 0x2e: // Nehalem EX
631 *Type
= X86::INTEL_COREI7
; // "nehalem"
632 *Subtype
= X86::INTEL_COREI7_NEHALEM
;
634 case 0x25: // Intel Core i7, laptop version.
635 case 0x2c: // Intel Core i7 processor and Intel Xeon processor. All
636 // processors are manufactured using the 32 nm process.
637 case 0x2f: // Westmere EX
638 *Type
= X86::INTEL_COREI7
; // "westmere"
639 *Subtype
= X86::INTEL_COREI7_WESTMERE
;
641 case 0x2a: // Intel Core i7 processor. All processors are manufactured
642 // using the 32 nm process.
644 *Type
= X86::INTEL_COREI7
; //"sandybridge"
645 *Subtype
= X86::INTEL_COREI7_SANDYBRIDGE
;
648 case 0x3e: // Ivy Bridge EP
649 *Type
= X86::INTEL_COREI7
; // "ivybridge"
650 *Subtype
= X86::INTEL_COREI7_IVYBRIDGE
;
658 *Type
= X86::INTEL_COREI7
; // "haswell"
659 *Subtype
= X86::INTEL_COREI7_HASWELL
;
667 *Type
= X86::INTEL_COREI7
; // "broadwell"
668 *Subtype
= X86::INTEL_COREI7_BROADWELL
;
672 case 0x4e: // Skylake mobile
673 case 0x5e: // Skylake desktop
674 case 0x8e: // Kaby Lake mobile
675 case 0x9e: // Kaby Lake desktop
676 *Type
= X86::INTEL_COREI7
; // "skylake"
677 *Subtype
= X86::INTEL_COREI7_SKYLAKE
;
682 *Type
= X86::INTEL_COREI7
;
683 if (Features2
& (1 << (X86::FEATURE_AVX512BF16
- 32)))
684 *Subtype
= X86::INTEL_COREI7_COOPERLAKE
; // "cooperlake"
685 else if (Features2
& (1 << (X86::FEATURE_AVX512VNNI
- 32)))
686 *Subtype
= X86::INTEL_COREI7_CASCADELAKE
; // "cascadelake"
688 *Subtype
= X86::INTEL_COREI7_SKYLAKE_AVX512
; // "skylake-avx512"
693 *Type
= X86::INTEL_COREI7
;
694 *Subtype
= X86::INTEL_COREI7_CANNONLAKE
; // "cannonlake"
700 *Type
= X86::INTEL_COREI7
;
701 *Subtype
= X86::INTEL_COREI7_ICELAKE_CLIENT
; // "icelake-client"
707 *Type
= X86::INTEL_COREI7
;
708 *Subtype
= X86::INTEL_COREI7_ICELAKE_SERVER
; // "icelake-server"
711 case 0x1c: // Most 45 nm Intel Atom processors
712 case 0x26: // 45 nm Atom Lincroft
713 case 0x27: // 32 nm Atom Medfield
714 case 0x35: // 32 nm Atom Midview
715 case 0x36: // 32 nm Atom Midview
716 *Type
= X86::INTEL_BONNELL
;
719 // Atom Silvermont codes from the Intel software optimization guide.
725 case 0x4c: // really airmont
726 *Type
= X86::INTEL_SILVERMONT
;
727 break; // "silvermont"
729 case 0x5c: // Apollo Lake
730 case 0x5f: // Denverton
731 *Type
= X86::INTEL_GOLDMONT
;
734 *Type
= X86::INTEL_GOLDMONT_PLUS
;
737 *Type
= X86::INTEL_TREMONT
;
741 *Type
= X86::INTEL_KNL
; // knl
745 *Type
= X86::INTEL_KNM
; // knm
748 default: // Unknown family 6 CPU, try to guess.
749 // TODO detect tigerlake host
750 if (Features3
& (1 << (X86::FEATURE_AVX512VP2INTERSECT
- 64))) {
751 *Type
= X86::INTEL_COREI7
;
752 *Subtype
= X86::INTEL_COREI7_TIGERLAKE
;
756 if (Features
& (1 << X86::FEATURE_AVX512VBMI2
)) {
757 *Type
= X86::INTEL_COREI7
;
758 *Subtype
= X86::INTEL_COREI7_ICELAKE_CLIENT
;
762 if (Features
& (1 << X86::FEATURE_AVX512VBMI
)) {
763 *Type
= X86::INTEL_COREI7
;
764 *Subtype
= X86::INTEL_COREI7_CANNONLAKE
;
768 if (Features2
& (1 << (X86::FEATURE_AVX512BF16
- 32))) {
769 *Type
= X86::INTEL_COREI7
;
770 *Subtype
= X86::INTEL_COREI7_COOPERLAKE
;
774 if (Features2
& (1 << (X86::FEATURE_AVX512VNNI
- 32))) {
775 *Type
= X86::INTEL_COREI7
;
776 *Subtype
= X86::INTEL_COREI7_CASCADELAKE
;
780 if (Features
& (1 << X86::FEATURE_AVX512VL
)) {
781 *Type
= X86::INTEL_COREI7
;
782 *Subtype
= X86::INTEL_COREI7_SKYLAKE_AVX512
;
786 if (Features
& (1 << X86::FEATURE_AVX512ER
)) {
787 *Type
= X86::INTEL_KNL
; // knl
791 if (Features3
& (1 << (X86::FEATURE_CLFLUSHOPT
- 64))) {
792 if (Features3
& (1 << (X86::FEATURE_SHA
- 64))) {
793 *Type
= X86::INTEL_GOLDMONT
;
795 *Type
= X86::INTEL_COREI7
;
796 *Subtype
= X86::INTEL_COREI7_SKYLAKE
;
800 if (Features3
& (1 << (X86::FEATURE_ADX
- 64))) {
801 *Type
= X86::INTEL_COREI7
;
802 *Subtype
= X86::INTEL_COREI7_BROADWELL
;
805 if (Features
& (1 << X86::FEATURE_AVX2
)) {
806 *Type
= X86::INTEL_COREI7
;
807 *Subtype
= X86::INTEL_COREI7_HASWELL
;
810 if (Features
& (1 << X86::FEATURE_AVX
)) {
811 *Type
= X86::INTEL_COREI7
;
812 *Subtype
= X86::INTEL_COREI7_SANDYBRIDGE
;
815 if (Features
& (1 << X86::FEATURE_SSE4_2
)) {
816 if (Features3
& (1 << (X86::FEATURE_MOVBE
- 64))) {
817 *Type
= X86::INTEL_SILVERMONT
;
819 *Type
= X86::INTEL_COREI7
;
820 *Subtype
= X86::INTEL_COREI7_NEHALEM
;
824 if (Features
& (1 << X86::FEATURE_SSE4_1
)) {
825 *Type
= X86::INTEL_CORE2
; // "penryn"
826 *Subtype
= X86::INTEL_CORE2_45
;
829 if (Features
& (1 << X86::FEATURE_SSSE3
)) {
830 if (Features3
& (1 << (X86::FEATURE_MOVBE
- 64))) {
831 *Type
= X86::INTEL_BONNELL
; // "bonnell"
833 *Type
= X86::INTEL_CORE2
; // "core2"
834 *Subtype
= X86::INTEL_CORE2_65
;
838 if (Features3
& (1 << (X86::FEATURE_EM64T
- 64))) {
839 *Type
= X86::INTEL_CORE2
; // "core2"
840 *Subtype
= X86::INTEL_CORE2_65
;
843 if (Features
& (1 << X86::FEATURE_SSE3
)) {
844 *Type
= X86::INTEL_CORE_DUO
;
847 if (Features
& (1 << X86::FEATURE_SSE2
)) {
848 *Type
= X86::INTEL_PENTIUM_M
;
851 if (Features
& (1 << X86::FEATURE_SSE
)) {
852 *Type
= X86::INTEL_PENTIUM_III
;
855 if (Features
& (1 << X86::FEATURE_MMX
)) {
856 *Type
= X86::INTEL_PENTIUM_II
;
859 *Type
= X86::INTEL_PENTIUM_PRO
;
864 if (Features3
& (1 << (X86::FEATURE_EM64T
- 64))) {
865 *Type
= X86::INTEL_NOCONA
;
868 if (Features
& (1 << X86::FEATURE_SSE3
)) {
869 *Type
= X86::INTEL_PRESCOTT
;
872 *Type
= X86::INTEL_PENTIUM_IV
;
880 static void getAMDProcessorTypeAndSubtype(unsigned Family
, unsigned Model
,
881 unsigned Features
, unsigned *Type
,
883 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
884 // appears to be no way to generate the wide variety of AMD-specific targets
885 // from the information returned from CPUID.
888 *Type
= X86::AMD_i486
;
891 *Type
= X86::AMDPENTIUM
;
895 *Subtype
= X86::AMDPENTIUM_K6
;
898 *Subtype
= X86::AMDPENTIUM_K62
;
902 *Subtype
= X86::AMDPENTIUM_K63
;
905 *Subtype
= X86::AMDPENTIUM_GEODE
;
910 if (Features
& (1 << X86::FEATURE_SSE
)) {
911 *Type
= X86::AMD_ATHLON_XP
;
912 break; // "athlon-xp"
914 *Type
= X86::AMD_ATHLON
;
917 if (Features
& (1 << X86::FEATURE_SSE3
)) {
918 *Type
= X86::AMD_K8SSE3
;
924 *Type
= X86::AMDFAM10H
; // "amdfam10"
927 *Subtype
= X86::AMDFAM10H_BARCELONA
;
930 *Subtype
= X86::AMDFAM10H_SHANGHAI
;
933 *Subtype
= X86::AMDFAM10H_ISTANBUL
;
938 *Type
= X86::AMD_BTVER1
;
941 *Type
= X86::AMDFAM15H
;
942 if (Model
>= 0x60 && Model
<= 0x7f) {
943 *Subtype
= X86::AMDFAM15H_BDVER4
;
944 break; // "bdver4"; 60h-7Fh: Excavator
946 if (Model
>= 0x30 && Model
<= 0x3f) {
947 *Subtype
= X86::AMDFAM15H_BDVER3
;
948 break; // "bdver3"; 30h-3Fh: Steamroller
950 if ((Model
>= 0x10 && Model
<= 0x1f) || Model
== 0x02) {
951 *Subtype
= X86::AMDFAM15H_BDVER2
;
952 break; // "bdver2"; 02h, 10h-1Fh: Piledriver
955 *Subtype
= X86::AMDFAM15H_BDVER1
;
956 break; // "bdver1"; 00h-0Fh: Bulldozer
960 *Type
= X86::AMD_BTVER2
;
963 *Type
= X86::AMDFAM17H
;
964 if (Model
>= 0x30 && Model
<= 0x3f) {
965 *Subtype
= X86::AMDFAM17H_ZNVER2
;
966 break; // "znver2"; 30h-3fh: Zen2
969 *Subtype
= X86::AMDFAM17H_ZNVER1
;
970 break; // "znver1"; 00h-0Fh: Zen1
978 static void getAvailableFeatures(unsigned ECX
, unsigned EDX
, unsigned MaxLeaf
,
979 unsigned *FeaturesOut
, unsigned *Features2Out
,
980 unsigned *Features3Out
) {
981 unsigned Features
= 0;
982 unsigned Features2
= 0;
983 unsigned Features3
= 0;
986 auto setFeature
= [&](unsigned F
) {
988 Features
|= 1U << (F
& 0x1f);
990 Features2
|= 1U << ((F
- 32) & 0x1f);
992 Features3
|= 1U << ((F
- 64) & 0x1f);
994 llvm_unreachable("Unexpected FeatureBit");
998 setFeature(X86::FEATURE_CMOV
);
1000 setFeature(X86::FEATURE_MMX
);
1001 if ((EDX
>> 25) & 1)
1002 setFeature(X86::FEATURE_SSE
);
1003 if ((EDX
>> 26) & 1)
1004 setFeature(X86::FEATURE_SSE2
);
1007 setFeature(X86::FEATURE_SSE3
);
1009 setFeature(X86::FEATURE_PCLMUL
);
1011 setFeature(X86::FEATURE_SSSE3
);
1012 if ((ECX
>> 12) & 1)
1013 setFeature(X86::FEATURE_FMA
);
1014 if ((ECX
>> 19) & 1)
1015 setFeature(X86::FEATURE_SSE4_1
);
1016 if ((ECX
>> 20) & 1)
1017 setFeature(X86::FEATURE_SSE4_2
);
1018 if ((ECX
>> 23) & 1)
1019 setFeature(X86::FEATURE_POPCNT
);
1020 if ((ECX
>> 25) & 1)
1021 setFeature(X86::FEATURE_AES
);
1023 if ((ECX
>> 22) & 1)
1024 setFeature(X86::FEATURE_MOVBE
);
1026 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
1027 // indicates that the AVX registers will be saved and restored on context
1028 // switch, then we have full AVX support.
1029 const unsigned AVXBits
= (1 << 27) | (1 << 28);
1030 bool HasAVX
= ((ECX
& AVXBits
) == AVXBits
) && !getX86XCR0(&EAX
, &EDX
) &&
1031 ((EAX
& 0x6) == 0x6);
1032 bool HasAVX512Save
= HasAVX
&& ((EAX
& 0xe0) == 0xe0);
1035 setFeature(X86::FEATURE_AVX
);
1038 MaxLeaf
>= 0x7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX
, &EBX
, &ECX
, &EDX
);
1040 if (HasLeaf7
&& ((EBX
>> 3) & 1))
1041 setFeature(X86::FEATURE_BMI
);
1042 if (HasLeaf7
&& ((EBX
>> 5) & 1) && HasAVX
)
1043 setFeature(X86::FEATURE_AVX2
);
1044 if (HasLeaf7
&& ((EBX
>> 8) & 1))
1045 setFeature(X86::FEATURE_BMI2
);
1046 if (HasLeaf7
&& ((EBX
>> 16) & 1) && HasAVX512Save
)
1047 setFeature(X86::FEATURE_AVX512F
);
1048 if (HasLeaf7
&& ((EBX
>> 17) & 1) && HasAVX512Save
)
1049 setFeature(X86::FEATURE_AVX512DQ
);
1050 if (HasLeaf7
&& ((EBX
>> 19) & 1))
1051 setFeature(X86::FEATURE_ADX
);
1052 if (HasLeaf7
&& ((EBX
>> 21) & 1) && HasAVX512Save
)
1053 setFeature(X86::FEATURE_AVX512IFMA
);
1054 if (HasLeaf7
&& ((EBX
>> 23) & 1))
1055 setFeature(X86::FEATURE_CLFLUSHOPT
);
1056 if (HasLeaf7
&& ((EBX
>> 26) & 1) && HasAVX512Save
)
1057 setFeature(X86::FEATURE_AVX512PF
);
1058 if (HasLeaf7
&& ((EBX
>> 27) & 1) && HasAVX512Save
)
1059 setFeature(X86::FEATURE_AVX512ER
);
1060 if (HasLeaf7
&& ((EBX
>> 28) & 1) && HasAVX512Save
)
1061 setFeature(X86::FEATURE_AVX512CD
);
1062 if (HasLeaf7
&& ((EBX
>> 29) & 1))
1063 setFeature(X86::FEATURE_SHA
);
1064 if (HasLeaf7
&& ((EBX
>> 30) & 1) && HasAVX512Save
)
1065 setFeature(X86::FEATURE_AVX512BW
);
1066 if (HasLeaf7
&& ((EBX
>> 31) & 1) && HasAVX512Save
)
1067 setFeature(X86::FEATURE_AVX512VL
);
1069 if (HasLeaf7
&& ((ECX
>> 1) & 1) && HasAVX512Save
)
1070 setFeature(X86::FEATURE_AVX512VBMI
);
1071 if (HasLeaf7
&& ((ECX
>> 6) & 1) && HasAVX512Save
)
1072 setFeature(X86::FEATURE_AVX512VBMI2
);
1073 if (HasLeaf7
&& ((ECX
>> 8) & 1))
1074 setFeature(X86::FEATURE_GFNI
);
1075 if (HasLeaf7
&& ((ECX
>> 10) & 1) && HasAVX
)
1076 setFeature(X86::FEATURE_VPCLMULQDQ
);
1077 if (HasLeaf7
&& ((ECX
>> 11) & 1) && HasAVX512Save
)
1078 setFeature(X86::FEATURE_AVX512VNNI
);
1079 if (HasLeaf7
&& ((ECX
>> 12) & 1) && HasAVX512Save
)
1080 setFeature(X86::FEATURE_AVX512BITALG
);
1081 if (HasLeaf7
&& ((ECX
>> 14) & 1) && HasAVX512Save
)
1082 setFeature(X86::FEATURE_AVX512VPOPCNTDQ
);
1084 if (HasLeaf7
&& ((EDX
>> 2) & 1) && HasAVX512Save
)
1085 setFeature(X86::FEATURE_AVX5124VNNIW
);
1086 if (HasLeaf7
&& ((EDX
>> 3) & 1) && HasAVX512Save
)
1087 setFeature(X86::FEATURE_AVX5124FMAPS
);
1088 if (HasLeaf7
&& ((EDX
>> 8) & 1) && HasAVX512Save
)
1089 setFeature(X86::FEATURE_AVX512VP2INTERSECT
);
1091 bool HasLeaf7Subleaf1
=
1092 MaxLeaf
>= 7 && !getX86CpuIDAndInfoEx(0x7, 0x1, &EAX
, &EBX
, &ECX
, &EDX
);
1093 if (HasLeaf7Subleaf1
&& ((EAX
>> 5) & 1) && HasAVX512Save
)
1094 setFeature(X86::FEATURE_AVX512BF16
);
1096 unsigned MaxExtLevel
;
1097 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel
, &EBX
, &ECX
, &EDX
);
1099 bool HasExtLeaf1
= MaxExtLevel
>= 0x80000001 &&
1100 !getX86CpuIDAndInfo(0x80000001, &EAX
, &EBX
, &ECX
, &EDX
);
1101 if (HasExtLeaf1
&& ((ECX
>> 6) & 1))
1102 setFeature(X86::FEATURE_SSE4_A
);
1103 if (HasExtLeaf1
&& ((ECX
>> 11) & 1))
1104 setFeature(X86::FEATURE_XOP
);
1105 if (HasExtLeaf1
&& ((ECX
>> 16) & 1))
1106 setFeature(X86::FEATURE_FMA4
);
1108 if (HasExtLeaf1
&& ((EDX
>> 29) & 1))
1109 setFeature(X86::FEATURE_EM64T
);
1111 *FeaturesOut
= Features
;
1112 *Features2Out
= Features2
;
1113 *Features3Out
= Features3
;
1116 StringRef
sys::getHostCPUName() {
1117 unsigned EAX
= 0, EBX
= 0, ECX
= 0, EDX
= 0;
1118 unsigned MaxLeaf
, Vendor
;
1120 #if defined(__GNUC__) || defined(__clang__)
1121 //FIXME: include cpuid.h from clang or copy __get_cpuid_max here
1122 // and simplify it to not invoke __cpuid (like cpu_model.c in
1123 // compiler-rt/lib/builtins/cpu_model.c?
1124 // Opting for the second option.
1125 if(!isCpuIdSupported())
1128 if (getX86CpuIDAndInfo(0, &MaxLeaf
, &Vendor
, &ECX
, &EDX
) || MaxLeaf
< 1)
1130 getX86CpuIDAndInfo(0x1, &EAX
, &EBX
, &ECX
, &EDX
);
1132 unsigned Brand_id
= EBX
& 0xff;
1133 unsigned Family
= 0, Model
= 0;
1134 unsigned Features
= 0, Features2
= 0, Features3
= 0;
1135 detectX86FamilyModel(EAX
, &Family
, &Model
);
1136 getAvailableFeatures(ECX
, EDX
, MaxLeaf
, &Features
, &Features2
, &Features3
);
1139 unsigned Subtype
= 0;
1141 if (Vendor
== SIG_INTEL
) {
1142 getIntelProcessorTypeAndSubtype(Family
, Model
, Brand_id
, Features
,
1143 Features2
, Features3
, &Type
, &Subtype
);
1144 } else if (Vendor
== SIG_AMD
) {
1145 getAMDProcessorTypeAndSubtype(Family
, Model
, Features
, &Type
, &Subtype
);
1148 // Check subtypes first since those are more specific.
1149 #define X86_CPU_SUBTYPE(ARCHNAME, ENUM) \
1150 if (Subtype == X86::ENUM) \
1152 #include "llvm/Support/X86TargetParser.def"
1155 #define X86_CPU_TYPE(ARCHNAME, ENUM) \
1156 if (Type == X86::ENUM) \
1158 #include "llvm/Support/X86TargetParser.def"
1163 #elif defined(__APPLE__) && (defined(__ppc__) || defined(__powerpc__))
1164 StringRef
sys::getHostCPUName() {
1165 host_basic_info_data_t hostInfo
;
1166 mach_msg_type_number_t infoCount
;
1168 infoCount
= HOST_BASIC_INFO_COUNT
;
1169 mach_port_t hostPort
= mach_host_self();
1170 host_info(hostPort
, HOST_BASIC_INFO
, (host_info_t
)&hostInfo
,
1172 mach_port_deallocate(mach_task_self(), hostPort
);
1174 if (hostInfo
.cpu_type
!= CPU_TYPE_POWERPC
)
1177 switch (hostInfo
.cpu_subtype
) {
1178 case CPU_SUBTYPE_POWERPC_601
:
1180 case CPU_SUBTYPE_POWERPC_602
:
1182 case CPU_SUBTYPE_POWERPC_603
:
1184 case CPU_SUBTYPE_POWERPC_603e
:
1186 case CPU_SUBTYPE_POWERPC_603ev
:
1188 case CPU_SUBTYPE_POWERPC_604
:
1190 case CPU_SUBTYPE_POWERPC_604e
:
1192 case CPU_SUBTYPE_POWERPC_620
:
1194 case CPU_SUBTYPE_POWERPC_750
:
1196 case CPU_SUBTYPE_POWERPC_7400
:
1198 case CPU_SUBTYPE_POWERPC_7450
:
1200 case CPU_SUBTYPE_POWERPC_970
:
1207 #elif defined(__linux__) && (defined(__ppc__) || defined(__powerpc__))
1208 StringRef
sys::getHostCPUName() {
1209 std::unique_ptr
<llvm::MemoryBuffer
> P
= getProcCpuinfoContent();
1210 StringRef Content
= P
? P
->getBuffer() : "";
1211 return detail::getHostCPUNameForPowerPC(Content
);
1213 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
1214 StringRef
sys::getHostCPUName() {
1215 std::unique_ptr
<llvm::MemoryBuffer
> P
= getProcCpuinfoContent();
1216 StringRef Content
= P
? P
->getBuffer() : "";
1217 return detail::getHostCPUNameForARM(Content
);
1219 #elif defined(__linux__) && defined(__s390x__)
1220 StringRef
sys::getHostCPUName() {
1221 std::unique_ptr
<llvm::MemoryBuffer
> P
= getProcCpuinfoContent();
1222 StringRef Content
= P
? P
->getBuffer() : "";
1223 return detail::getHostCPUNameForS390x(Content
);
1226 StringRef
sys::getHostCPUName() { return "generic"; }
1229 #if defined(__linux__) && defined(__x86_64__)
1230 // On Linux, the number of physical cores can be computed from /proc/cpuinfo,
1231 // using the number of unique physical/core id pairs. The following
1232 // implementation reads the /proc/cpuinfo format on an x86_64 system.
1233 static int computeHostNumPhysicalCores() {
1234 // Read /proc/cpuinfo as a stream (until EOF reached). It cannot be
1235 // mmapped because it appears to have 0 size.
1236 llvm::ErrorOr
<std::unique_ptr
<llvm::MemoryBuffer
>> Text
=
1237 llvm::MemoryBuffer::getFileAsStream("/proc/cpuinfo");
1238 if (std::error_code EC
= Text
.getError()) {
1239 llvm::errs() << "Can't read "
1240 << "/proc/cpuinfo: " << EC
.message() << "\n";
1243 SmallVector
<StringRef
, 8> strs
;
1244 (*Text
)->getBuffer().split(strs
, "\n", /*MaxSplit=*/-1,
1245 /*KeepEmpty=*/false);
1246 int CurPhysicalId
= -1;
1248 SmallSet
<std::pair
<int, int>, 32> UniqueItems
;
1249 for (auto &Line
: strs
) {
1251 if (!Line
.startswith("physical id") && !Line
.startswith("core id"))
1253 std::pair
<StringRef
, StringRef
> Data
= Line
.split(':');
1254 auto Name
= Data
.first
.trim();
1255 auto Val
= Data
.second
.trim();
1256 if (Name
== "physical id") {
1257 assert(CurPhysicalId
== -1 &&
1258 "Expected a core id before seeing another physical id");
1259 Val
.getAsInteger(10, CurPhysicalId
);
1261 if (Name
== "core id") {
1262 assert(CurCoreId
== -1 &&
1263 "Expected a physical id before seeing another core id");
1264 Val
.getAsInteger(10, CurCoreId
);
1266 if (CurPhysicalId
!= -1 && CurCoreId
!= -1) {
1267 UniqueItems
.insert(std::make_pair(CurPhysicalId
, CurCoreId
));
1272 return UniqueItems
.size();
1274 #elif defined(__APPLE__) && defined(__x86_64__)
1275 #include <sys/param.h>
1276 #include <sys/sysctl.h>
1278 // Gets the number of *physical cores* on the machine.
1279 static int computeHostNumPhysicalCores() {
1281 size_t len
= sizeof(count
);
1282 sysctlbyname("hw.physicalcpu", &count
, &len
, NULL
, 0);
1286 nm
[1] = HW_AVAILCPU
;
1287 sysctl(nm
, 2, &count
, &len
, NULL
, 0);
1294 // On other systems, return -1 to indicate unknown.
1295 static int computeHostNumPhysicalCores() { return -1; }
1298 int sys::getHostNumPhysicalCores() {
1299 static int NumCores
= computeHostNumPhysicalCores();
1303 #if defined(__i386__) || defined(_M_IX86) || \
1304 defined(__x86_64__) || defined(_M_X64)
1305 bool sys::getHostCPUFeatures(StringMap
<bool> &Features
) {
1306 unsigned EAX
= 0, EBX
= 0, ECX
= 0, EDX
= 0;
1313 if (getX86CpuIDAndInfo(0, &MaxLevel
, text
.u
+ 0, text
.u
+ 2, text
.u
+ 1) ||
1317 getX86CpuIDAndInfo(1, &EAX
, &EBX
, &ECX
, &EDX
);
1319 Features
["cx8"] = (EDX
>> 8) & 1;
1320 Features
["cmov"] = (EDX
>> 15) & 1;
1321 Features
["mmx"] = (EDX
>> 23) & 1;
1322 Features
["fxsr"] = (EDX
>> 24) & 1;
1323 Features
["sse"] = (EDX
>> 25) & 1;
1324 Features
["sse2"] = (EDX
>> 26) & 1;
1326 Features
["sse3"] = (ECX
>> 0) & 1;
1327 Features
["pclmul"] = (ECX
>> 1) & 1;
1328 Features
["ssse3"] = (ECX
>> 9) & 1;
1329 Features
["cx16"] = (ECX
>> 13) & 1;
1330 Features
["sse4.1"] = (ECX
>> 19) & 1;
1331 Features
["sse4.2"] = (ECX
>> 20) & 1;
1332 Features
["movbe"] = (ECX
>> 22) & 1;
1333 Features
["popcnt"] = (ECX
>> 23) & 1;
1334 Features
["aes"] = (ECX
>> 25) & 1;
1335 Features
["rdrnd"] = (ECX
>> 30) & 1;
1337 // If CPUID indicates support for XSAVE, XRESTORE and AVX, and XGETBV
1338 // indicates that the AVX registers will be saved and restored on context
1339 // switch, then we have full AVX support.
1340 bool HasAVXSave
= ((ECX
>> 27) & 1) && ((ECX
>> 28) & 1) &&
1341 !getX86XCR0(&EAX
, &EDX
) && ((EAX
& 0x6) == 0x6);
1342 // AVX512 requires additional context to be saved by the OS.
1343 bool HasAVX512Save
= HasAVXSave
&& ((EAX
& 0xe0) == 0xe0);
1345 Features
["avx"] = HasAVXSave
;
1346 Features
["fma"] = ((ECX
>> 12) & 1) && HasAVXSave
;
1347 // Only enable XSAVE if OS has enabled support for saving YMM state.
1348 Features
["xsave"] = ((ECX
>> 26) & 1) && HasAVXSave
;
1349 Features
["f16c"] = ((ECX
>> 29) & 1) && HasAVXSave
;
1351 unsigned MaxExtLevel
;
1352 getX86CpuIDAndInfo(0x80000000, &MaxExtLevel
, &EBX
, &ECX
, &EDX
);
1354 bool HasExtLeaf1
= MaxExtLevel
>= 0x80000001 &&
1355 !getX86CpuIDAndInfo(0x80000001, &EAX
, &EBX
, &ECX
, &EDX
);
1356 Features
["sahf"] = HasExtLeaf1
&& ((ECX
>> 0) & 1);
1357 Features
["lzcnt"] = HasExtLeaf1
&& ((ECX
>> 5) & 1);
1358 Features
["sse4a"] = HasExtLeaf1
&& ((ECX
>> 6) & 1);
1359 Features
["prfchw"] = HasExtLeaf1
&& ((ECX
>> 8) & 1);
1360 Features
["xop"] = HasExtLeaf1
&& ((ECX
>> 11) & 1) && HasAVXSave
;
1361 Features
["lwp"] = HasExtLeaf1
&& ((ECX
>> 15) & 1);
1362 Features
["fma4"] = HasExtLeaf1
&& ((ECX
>> 16) & 1) && HasAVXSave
;
1363 Features
["tbm"] = HasExtLeaf1
&& ((ECX
>> 21) & 1);
1364 Features
["mwaitx"] = HasExtLeaf1
&& ((ECX
>> 29) & 1);
1366 Features
["64bit"] = HasExtLeaf1
&& ((EDX
>> 29) & 1);
1368 // Miscellaneous memory related features, detected by
1369 // using the 0x80000008 leaf of the CPUID instruction
1370 bool HasExtLeaf8
= MaxExtLevel
>= 0x80000008 &&
1371 !getX86CpuIDAndInfo(0x80000008, &EAX
, &EBX
, &ECX
, &EDX
);
1372 Features
["clzero"] = HasExtLeaf8
&& ((EBX
>> 0) & 1);
1373 Features
["wbnoinvd"] = HasExtLeaf8
&& ((EBX
>> 9) & 1);
1376 MaxLevel
>= 7 && !getX86CpuIDAndInfoEx(0x7, 0x0, &EAX
, &EBX
, &ECX
, &EDX
);
1378 Features
["fsgsbase"] = HasLeaf7
&& ((EBX
>> 0) & 1);
1379 Features
["sgx"] = HasLeaf7
&& ((EBX
>> 2) & 1);
1380 Features
["bmi"] = HasLeaf7
&& ((EBX
>> 3) & 1);
1381 // AVX2 is only supported if we have the OS save support from AVX.
1382 Features
["avx2"] = HasLeaf7
&& ((EBX
>> 5) & 1) && HasAVXSave
;
1383 Features
["bmi2"] = HasLeaf7
&& ((EBX
>> 8) & 1);
1384 Features
["invpcid"] = HasLeaf7
&& ((EBX
>> 10) & 1);
1385 Features
["rtm"] = HasLeaf7
&& ((EBX
>> 11) & 1);
1386 // AVX512 is only supported if the OS supports the context save for it.
1387 Features
["avx512f"] = HasLeaf7
&& ((EBX
>> 16) & 1) && HasAVX512Save
;
1388 Features
["avx512dq"] = HasLeaf7
&& ((EBX
>> 17) & 1) && HasAVX512Save
;
1389 Features
["rdseed"] = HasLeaf7
&& ((EBX
>> 18) & 1);
1390 Features
["adx"] = HasLeaf7
&& ((EBX
>> 19) & 1);
1391 Features
["avx512ifma"] = HasLeaf7
&& ((EBX
>> 21) & 1) && HasAVX512Save
;
1392 Features
["clflushopt"] = HasLeaf7
&& ((EBX
>> 23) & 1);
1393 Features
["clwb"] = HasLeaf7
&& ((EBX
>> 24) & 1);
1394 Features
["avx512pf"] = HasLeaf7
&& ((EBX
>> 26) & 1) && HasAVX512Save
;
1395 Features
["avx512er"] = HasLeaf7
&& ((EBX
>> 27) & 1) && HasAVX512Save
;
1396 Features
["avx512cd"] = HasLeaf7
&& ((EBX
>> 28) & 1) && HasAVX512Save
;
1397 Features
["sha"] = HasLeaf7
&& ((EBX
>> 29) & 1);
1398 Features
["avx512bw"] = HasLeaf7
&& ((EBX
>> 30) & 1) && HasAVX512Save
;
1399 Features
["avx512vl"] = HasLeaf7
&& ((EBX
>> 31) & 1) && HasAVX512Save
;
1401 Features
["prefetchwt1"] = HasLeaf7
&& ((ECX
>> 0) & 1);
1402 Features
["avx512vbmi"] = HasLeaf7
&& ((ECX
>> 1) & 1) && HasAVX512Save
;
1403 Features
["pku"] = HasLeaf7
&& ((ECX
>> 4) & 1);
1404 Features
["waitpkg"] = HasLeaf7
&& ((ECX
>> 5) & 1);
1405 Features
["avx512vbmi2"] = HasLeaf7
&& ((ECX
>> 6) & 1) && HasAVX512Save
;
1406 Features
["shstk"] = HasLeaf7
&& ((ECX
>> 7) & 1);
1407 Features
["gfni"] = HasLeaf7
&& ((ECX
>> 8) & 1);
1408 Features
["vaes"] = HasLeaf7
&& ((ECX
>> 9) & 1) && HasAVXSave
;
1409 Features
["vpclmulqdq"] = HasLeaf7
&& ((ECX
>> 10) & 1) && HasAVXSave
;
1410 Features
["avx512vnni"] = HasLeaf7
&& ((ECX
>> 11) & 1) && HasAVX512Save
;
1411 Features
["avx512bitalg"] = HasLeaf7
&& ((ECX
>> 12) & 1) && HasAVX512Save
;
1412 Features
["avx512vpopcntdq"] = HasLeaf7
&& ((ECX
>> 14) & 1) && HasAVX512Save
;
1413 Features
["rdpid"] = HasLeaf7
&& ((ECX
>> 22) & 1);
1414 Features
["cldemote"] = HasLeaf7
&& ((ECX
>> 25) & 1);
1415 Features
["movdiri"] = HasLeaf7
&& ((ECX
>> 27) & 1);
1416 Features
["movdir64b"] = HasLeaf7
&& ((ECX
>> 28) & 1);
1417 Features
["enqcmd"] = HasLeaf7
&& ((ECX
>> 29) & 1);
1419 // There are two CPUID leafs which information associated with the pconfig
1421 // EAX=0x7, ECX=0x0 indicates the availability of the instruction (via the 18th
1422 // bit of EDX), while the EAX=0x1b leaf returns information on the
1423 // availability of specific pconfig leafs.
1424 // The target feature here only refers to the the first of these two.
1425 // Users might need to check for the availability of specific pconfig
1426 // leaves using cpuid, since that information is ignored while
1427 // detecting features using the "-march=native" flag.
1428 // For more info, see X86 ISA docs.
1429 Features
["pconfig"] = HasLeaf7
&& ((EDX
>> 18) & 1);
1430 bool HasLeaf7Subleaf1
=
1431 MaxLevel
>= 7 && !getX86CpuIDAndInfoEx(0x7, 0x1, &EAX
, &EBX
, &ECX
, &EDX
);
1432 Features
["avx512bf16"] = HasLeaf7Subleaf1
&& ((EAX
>> 5) & 1) && HasAVX512Save
;
1434 bool HasLeafD
= MaxLevel
>= 0xd &&
1435 !getX86CpuIDAndInfoEx(0xd, 0x1, &EAX
, &EBX
, &ECX
, &EDX
);
1437 // Only enable XSAVE if OS has enabled support for saving YMM state.
1438 Features
["xsaveopt"] = HasLeafD
&& ((EAX
>> 0) & 1) && HasAVXSave
;
1439 Features
["xsavec"] = HasLeafD
&& ((EAX
>> 1) & 1) && HasAVXSave
;
1440 Features
["xsaves"] = HasLeafD
&& ((EAX
>> 3) & 1) && HasAVXSave
;
1442 bool HasLeaf14
= MaxLevel
>= 0x14 &&
1443 !getX86CpuIDAndInfoEx(0x14, 0x0, &EAX
, &EBX
, &ECX
, &EDX
);
1445 Features
["ptwrite"] = HasLeaf14
&& ((EBX
>> 4) & 1);
1449 #elif defined(__linux__) && (defined(__arm__) || defined(__aarch64__))
1450 bool sys::getHostCPUFeatures(StringMap
<bool> &Features
) {
1451 std::unique_ptr
<llvm::MemoryBuffer
> P
= getProcCpuinfoContent();
1455 SmallVector
<StringRef
, 32> Lines
;
1456 P
->getBuffer().split(Lines
, "\n");
1458 SmallVector
<StringRef
, 32> CPUFeatures
;
1460 // Look for the CPU features.
1461 for (unsigned I
= 0, E
= Lines
.size(); I
!= E
; ++I
)
1462 if (Lines
[I
].startswith("Features")) {
1463 Lines
[I
].split(CPUFeatures
, ' ');
1467 #if defined(__aarch64__)
1468 // Keep track of which crypto features we have seen
1469 enum { CAP_AES
= 0x1, CAP_PMULL
= 0x2, CAP_SHA1
= 0x4, CAP_SHA2
= 0x8 };
1470 uint32_t crypto
= 0;
1473 for (unsigned I
= 0, E
= CPUFeatures
.size(); I
!= E
; ++I
) {
1474 StringRef LLVMFeatureStr
= StringSwitch
<StringRef
>(CPUFeatures
[I
])
1475 #if defined(__aarch64__)
1476 .Case("asimd", "neon")
1477 .Case("fp", "fp-armv8")
1478 .Case("crc32", "crc")
1480 .Case("half", "fp16")
1481 .Case("neon", "neon")
1482 .Case("vfpv3", "vfp3")
1483 .Case("vfpv3d16", "d16")
1484 .Case("vfpv4", "vfp4")
1485 .Case("idiva", "hwdiv-arm")
1486 .Case("idivt", "hwdiv")
1490 #if defined(__aarch64__)
1491 // We need to check crypto separately since we need all of the crypto
1492 // extensions to enable the subtarget feature
1493 if (CPUFeatures
[I
] == "aes")
1495 else if (CPUFeatures
[I
] == "pmull")
1496 crypto
|= CAP_PMULL
;
1497 else if (CPUFeatures
[I
] == "sha1")
1499 else if (CPUFeatures
[I
] == "sha2")
1503 if (LLVMFeatureStr
!= "")
1504 Features
[LLVMFeatureStr
] = true;
1507 #if defined(__aarch64__)
1508 // If we have all crypto bits we can add the feature
1509 if (crypto
== (CAP_AES
| CAP_PMULL
| CAP_SHA1
| CAP_SHA2
))
1510 Features
["crypto"] = true;
1515 #elif defined(_WIN32) && (defined(__aarch64__) || defined(_M_ARM64))
1516 bool sys::getHostCPUFeatures(StringMap
<bool> &Features
) {
1517 if (IsProcessorFeaturePresent(PF_ARM_NEON_INSTRUCTIONS_AVAILABLE
))
1518 Features
["neon"] = true;
1519 if (IsProcessorFeaturePresent(PF_ARM_V8_CRC32_INSTRUCTIONS_AVAILABLE
))
1520 Features
["crc"] = true;
1521 if (IsProcessorFeaturePresent(PF_ARM_V8_CRYPTO_INSTRUCTIONS_AVAILABLE
))
1522 Features
["crypto"] = true;
1527 bool sys::getHostCPUFeatures(StringMap
<bool> &Features
) { return false; }
1530 std::string
sys::getProcessTriple() {
1531 std::string TargetTripleString
= updateTripleOSVersion(LLVM_HOST_TRIPLE
);
1532 Triple
PT(Triple::normalize(TargetTripleString
));
1534 if (sizeof(void *) == 8 && PT
.isArch32Bit())
1535 PT
= PT
.get64BitArchVariant();
1536 if (sizeof(void *) == 4 && PT
.isArch64Bit())
1537 PT
= PT
.get32BitArchVariant();