cpu_x86.go

Documentation: internal/cpu

     1  // Copyright 2017 The Go Authors. All rights reserved.
     2  // Use of this source code is governed by a BSD-style
     3  // license that can be found in the LICENSE file.
     4  
     5  //go:build 386 || amd64
     6  
     7  package cpu
     8  
     9  const CacheLinePadSize = 64
    10  
    11  // cpuid is implemented in cpu_x86.s.
    12  func cpuid(eaxArg, ecxArg uint32) (eax, ebx, ecx, edx uint32)
    13  
    14  // xgetbv with ecx = 0 is implemented in cpu_x86.s.
    15  func xgetbv() (eax, edx uint32)
    16  
    17  // getGOAMD64level is implemented in cpu_x86.s. Returns number in [1,4].
    18  func getGOAMD64level() int32
    19  
    20  const (
    21  	// edx bits
    22  	cpuid_SSE2 = 1 << 26
    23  
    24  	// ecx bits
    25  	cpuid_SSE3      = 1 << 0
    26  	cpuid_PCLMULQDQ = 1 << 1
    27  	cpuid_SSSE3     = 1 << 9
    28  	cpuid_FMA       = 1 << 12
    29  	cpuid_SSE41     = 1 << 19
    30  	cpuid_SSE42     = 1 << 20
    31  	cpuid_POPCNT    = 1 << 23
    32  	cpuid_AES       = 1 << 25
    33  	cpuid_OSXSAVE   = 1 << 27
    34  	cpuid_AVX       = 1 << 28
    35  
    36  	// ebx bits
    37  	cpuid_BMI1     = 1 << 3
    38  	cpuid_AVX2     = 1 << 5
    39  	cpuid_BMI2     = 1 << 8
    40  	cpuid_ERMS     = 1 << 9
    41  	cpuid_AVX512F  = 1 << 16
    42  	cpuid_ADX      = 1 << 19
    43  	cpuid_SHA      = 1 << 29
    44  	cpuid_AVX512BW = 1 << 30
    45  	cpuid_AVX512VL = 1 << 31
    46  
    47  	// edx bits for CPUID 0x80000001
    48  	cpuid_RDTSCP = 1 << 27
    49  )
    50  
    51  var maxExtendedFunctionInformation uint32
    52  
    53  func doinit() {
    54  	options = []option{
    55  		{Name: "adx", Feature: &X86.HasADX},
    56  		{Name: "aes", Feature: &X86.HasAES},
    57  		{Name: "erms", Feature: &X86.HasERMS},
    58  		{Name: "pclmulqdq", Feature: &X86.HasPCLMULQDQ},
    59  		{Name: "rdtscp", Feature: &X86.HasRDTSCP},
    60  		{Name: "sha", Feature: &X86.HasSHA},
    61  	}
    62  	level := getGOAMD64level()
    63  	if level < 2 {
    64  		// These options are required at level 2. At lower levels
    65  		// they can be turned off.
    66  		options = append(options,
    67  			option{Name: "popcnt", Feature: &X86.HasPOPCNT},
    68  			option{Name: "sse3", Feature: &X86.HasSSE3},
    69  			option{Name: "sse41", Feature: &X86.HasSSE41},
    70  			option{Name: "sse42", Feature: &X86.HasSSE42},
    71  			option{Name: "ssse3", Feature: &X86.HasSSSE3})
    72  	}
    73  	if level < 3 {
    74  		// These options are required at level 3. At lower levels
    75  		// they can be turned off.
    76  		options = append(options,
    77  			option{Name: "avx", Feature: &X86.HasAVX},
    78  			option{Name: "avx2", Feature: &X86.HasAVX2},
    79  			option{Name: "bmi1", Feature: &X86.HasBMI1},
    80  			option{Name: "bmi2", Feature: &X86.HasBMI2},
    81  			option{Name: "fma", Feature: &X86.HasFMA})
    82  	}
    83  	if level < 4 {
    84  		// These options are required at level 4. At lower levels
    85  		// they can be turned off.
    86  		options = append(options,
    87  			option{Name: "avx512f", Feature: &X86.HasAVX512F},
    88  			option{Name: "avx512bw", Feature: &X86.HasAVX512BW},
    89  			option{Name: "avx512vl", Feature: &X86.HasAVX512VL},
    90  		)
    91  	}
    92  
    93  	maxID, _, _, _ := cpuid(0, 0)
    94  
    95  	if maxID < 1 {
    96  		return
    97  	}
    98  
    99  	maxExtendedFunctionInformation, _, _, _ = cpuid(0x80000000, 0)
   100  
   101  	_, _, ecx1, _ := cpuid(1, 0)
   102  
   103  	X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
   104  	X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
   105  	X86.HasSSSE3 = isSet(ecx1, cpuid_SSSE3)
   106  	X86.HasSSE41 = isSet(ecx1, cpuid_SSE41)
   107  	X86.HasSSE42 = isSet(ecx1, cpuid_SSE42)
   108  	X86.HasPOPCNT = isSet(ecx1, cpuid_POPCNT)
   109  	X86.HasAES = isSet(ecx1, cpuid_AES)
   110  
   111  	// OSXSAVE can be false when using older Operating Systems
   112  	// or when explicitly disabled on newer Operating Systems by
   113  	// e.g. setting the xsavedisable boot option on Windows 10.
   114  	X86.HasOSXSAVE = isSet(ecx1, cpuid_OSXSAVE)
   115  
   116  	// The FMA instruction set extension only has VEX prefixed instructions.
   117  	// VEX prefixed instructions require OSXSAVE to be enabled.
   118  	// See Intel 64 and IA-32 Architecture Software Developer’s Manual Volume 2
   119  	// Section 2.4 "AVX and SSE Instruction Exception Specification"
   120  	X86.HasFMA = isSet(ecx1, cpuid_FMA) && X86.HasOSXSAVE
   121  
   122  	osSupportsAVX := false
   123  	osSupportsAVX512 := false
   124  	// For XGETBV, OSXSAVE bit is required and sufficient.
   125  	if X86.HasOSXSAVE {
   126  		eax, _ := xgetbv()
   127  		// Check if XMM and YMM registers have OS support.
   128  		osSupportsAVX = isSet(eax, 1<<1) && isSet(eax, 1<<2)
   129  
   130  		// AVX512 detection does not work on Darwin,
   131  		// see https://github.com/golang/go/issues/49233
   132  		//
   133  		// Check if opmask, ZMMhi256 and Hi16_ZMM have OS support.
   134  		osSupportsAVX512 = osSupportsAVX && isSet(eax, 1<<5) && isSet(eax, 1<<6) && isSet(eax, 1<<7)
   135  	}
   136  
   137  	X86.HasAVX = isSet(ecx1, cpuid_AVX) && osSupportsAVX
   138  
   139  	if maxID < 7 {
   140  		return
   141  	}
   142  
   143  	_, ebx7, _, _ := cpuid(7, 0)
   144  	X86.HasBMI1 = isSet(ebx7, cpuid_BMI1)
   145  	X86.HasAVX2 = isSet(ebx7, cpuid_AVX2) && osSupportsAVX
   146  	X86.HasBMI2 = isSet(ebx7, cpuid_BMI2)
   147  	X86.HasERMS = isSet(ebx7, cpuid_ERMS)
   148  	X86.HasADX = isSet(ebx7, cpuid_ADX)
   149  	X86.HasSHA = isSet(ebx7, cpuid_SHA)
   150  
   151  	X86.HasAVX512F = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
   152  	if X86.HasAVX512F {
   153  		X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
   154  		X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
   155  	}
   156  
   157  	var maxExtendedInformation uint32
   158  	maxExtendedInformation, _, _, _ = cpuid(0x80000000, 0)
   159  
   160  	if maxExtendedInformation < 0x80000001 {
   161  		return
   162  	}
   163  
   164  	_, _, _, edxExt1 := cpuid(0x80000001, 0)
   165  	X86.HasRDTSCP = isSet(edxExt1, cpuid_RDTSCP)
   166  }
   167  
   168  func isSet(hwc uint32, value uint32) bool {
   169  	return hwc&value != 0
   170  }
   171  
   172  // Name returns the CPU name given by the vendor.
   173  // If the CPU name can not be determined an
   174  // empty string is returned.
   175  func Name() string {
   176  	if maxExtendedFunctionInformation < 0x80000004 {
   177  		return ""
   178  	}
   179  
   180  	data := make([]byte, 0, 3*4*4)
   181  
   182  	var eax, ebx, ecx, edx uint32
   183  	eax, ebx, ecx, edx = cpuid(0x80000002, 0)
   184  	data = appendBytes(data, eax, ebx, ecx, edx)
   185  	eax, ebx, ecx, edx = cpuid(0x80000003, 0)
   186  	data = appendBytes(data, eax, ebx, ecx, edx)
   187  	eax, ebx, ecx, edx = cpuid(0x80000004, 0)
   188  	data = appendBytes(data, eax, ebx, ecx, edx)
   189  
   190  	// Trim leading spaces.
   191  	for len(data) > 0 && data[0] == ' ' {
   192  		data = data[1:]
   193  	}
   194  
   195  	// Trim tail after and including the first null byte.
   196  	for i, c := range data {
   197  		if c == '\x00' {
   198  			data = data[:i]
   199  			break
   200  		}
   201  	}
   202  
   203  	return string(data)
   204  }
   205  
   206  func appendBytes(b []byte, args ...uint32) []byte {
   207  	for _, arg := range args {
   208  		b = append(b,
   209  			byte((arg >> 0)),
   210  			byte((arg >> 8)),
   211  			byte((arg >> 16)),
   212  			byte((arg >> 24)))
   213  	}
   214  	return b
   215  }
   216
View as plain text