cpu_x86.go

     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  	// ecx bits
    22  	cpuid_SSE3      = 1 << 0
    23  	cpuid_PCLMULQDQ = 1 << 1
    24  	cpuid_SSSE3     = 1 << 9
    25  	cpuid_FMA       = 1 << 12
    26  	cpuid_SSE41     = 1 << 19
    27  	cpuid_SSE42     = 1 << 20
    28  	cpuid_POPCNT    = 1 << 23
    29  	cpuid_AES       = 1 << 25
    30  	cpuid_OSXSAVE   = 1 << 27
    31  	cpuid_AVX       = 1 << 28
    32  
    33  	// ebx bits
    34  	cpuid_BMI1     = 1 << 3
    35  	cpuid_AVX2     = 1 << 5
    36  	cpuid_BMI2     = 1 << 8
    37  	cpuid_ERMS     = 1 << 9
    38  	cpuid_AVX512F  = 1 << 16
    39  	cpuid_AVX512DQ = 1 << 17
    40  	cpuid_ADX      = 1 << 19
    41  	cpuid_AVX512CD = 1 << 28
    42  	cpuid_SHA      = 1 << 29
    43  	cpuid_AVX512BW = 1 << 30
    44  	cpuid_AVX512VL = 1 << 31
    45  
    46  	// ecx bits
    47  	cpuid_AVX512VPCLMULQDQ = 1 << 10
    48  
    49  	// edx bits
    50  	cpuid_FSRM = 1 << 4
    51  	// edx bits for CPUID 0x80000001
    52  	cpuid_RDTSCP = 1 << 27
    53  )
    54  
    55  var maxExtendedFunctionInformation uint32
    56  
    57  func doinit() {
    58  	options = []option{
    59  		{Name: "adx", Feature: &X86.HasADX},
    60  		{Name: "aes", Feature: &X86.HasAES},
    61  		{Name: "erms", Feature: &X86.HasERMS},
    62  		{Name: "fsrm", Feature: &X86.HasFSRM},
    63  		{Name: "pclmulqdq", Feature: &X86.HasPCLMULQDQ},
    64  		{Name: "rdtscp", Feature: &X86.HasRDTSCP},
    65  		{Name: "sha", Feature: &X86.HasSHA},
    66  		{Name: "vpclmulqdq", Feature: &X86.HasAVX512VPCLMULQDQ},
    67  	}
    68  	level := getGOAMD64level()
    69  	if level < 2 {
    70  		// These options are required at level 2. At lower levels
    71  		// they can be turned off.
    72  		options = append(options,
    73  			option{Name: "popcnt", Feature: &X86.HasPOPCNT},
    74  			option{Name: "sse3", Feature: &X86.HasSSE3},
    75  			option{Name: "sse41", Feature: &X86.HasSSE41},
    76  			option{Name: "sse42", Feature: &X86.HasSSE42},
    77  			option{Name: "ssse3", Feature: &X86.HasSSSE3})
    78  	}
    79  	if level < 3 {
    80  		// These options are required at level 3. At lower levels
    81  		// they can be turned off.
    82  		options = append(options,
    83  			option{Name: "avx", Feature: &X86.HasAVX},
    84  			option{Name: "avx2", Feature: &X86.HasAVX2},
    85  			option{Name: "bmi1", Feature: &X86.HasBMI1},
    86  			option{Name: "bmi2", Feature: &X86.HasBMI2},
    87  			option{Name: "fma", Feature: &X86.HasFMA})
    88  	}
    89  	if level < 4 {
    90  		// These options are required at level 4. At lower levels
    91  		// they can be turned off.
    92  		options = append(options,
    93  			option{Name: "avx512f", Feature: &X86.HasAVX512F},
    94  			option{Name: "avx512cd", Feature: &X86.HasAVX512CD},
    95  			option{Name: "avx512bw", Feature: &X86.HasAVX512BW},
    96  			option{Name: "avx512dq", Feature: &X86.HasAVX512DQ},
    97  			option{Name: "avx512vl", Feature: &X86.HasAVX512VL},
    98  		)
    99  	}
   100  
   101  	maxID, _, _, _ := cpuid(0, 0)
   102  
   103  	if maxID < 1 {
   104  		return
   105  	}
   106  
   107  	maxExtendedFunctionInformation, _, _, _ = cpuid(0x80000000, 0)
   108  
   109  	_, _, ecx1, _ := cpuid(1, 0)
   110  
   111  	X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
   112  	X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
   113  	X86.HasSSSE3 = isSet(ecx1, cpuid_SSSE3)
   114  	X86.HasSSE41 = isSet(ecx1, cpuid_SSE41)
   115  	X86.HasSSE42 = isSet(ecx1, cpuid_SSE42)
   116  	X86.HasPOPCNT = isSet(ecx1, cpuid_POPCNT)
   117  	X86.HasAES = isSet(ecx1, cpuid_AES)
   118  
   119  	// OSXSAVE can be false when using older Operating Systems
   120  	// or when explicitly disabled on newer Operating Systems by
   121  	// e.g. setting the xsavedisable boot option on Windows 10.
   122  	X86.HasOSXSAVE = isSet(ecx1, cpuid_OSXSAVE)
   123  
   124  	// The FMA instruction set extension only has VEX prefixed instructions.
   125  	// VEX prefixed instructions require OSXSAVE to be enabled.
   126  	// See Intel 64 and IA-32 Architecture Software Developer’s Manual Volume 2
   127  	// Section 2.4 "AVX and SSE Instruction Exception Specification"
   128  	X86.HasFMA = isSet(ecx1, cpuid_FMA) && X86.HasOSXSAVE
   129  
   130  	osSupportsAVX := false
   131  	osSupportsAVX512 := false
   132  	// For XGETBV, OSXSAVE bit is required and sufficient.
   133  	if X86.HasOSXSAVE {
   134  		eax, _ := xgetbv()
   135  		// Check if XMM and YMM registers have OS support.
   136  		osSupportsAVX = isSet(eax, 1<<1) && isSet(eax, 1<<2)
   137  
   138  		// AVX512 detection does not work on Darwin,
   139  		// see https://github.com/golang/go/issues/49233
   140  		//
   141  		// Check if opmask, ZMMhi256 and Hi16_ZMM have OS support.
   142  		osSupportsAVX512 = osSupportsAVX && isSet(eax, 1<<5) && isSet(eax, 1<<6) && isSet(eax, 1<<7)
   143  	}
   144  
   145  	X86.HasAVX = isSet(ecx1, cpuid_AVX) && osSupportsAVX
   146  
   147  	if maxID < 7 {
   148  		return
   149  	}
   150  
   151  	_, ebx7, ecx7, edx7 := cpuid(7, 0)
   152  	X86.HasBMI1 = isSet(ebx7, cpuid_BMI1)
   153  	X86.HasAVX2 = isSet(ebx7, cpuid_AVX2) && osSupportsAVX
   154  	X86.HasBMI2 = isSet(ebx7, cpuid_BMI2)
   155  	X86.HasERMS = isSet(ebx7, cpuid_ERMS)
   156  	X86.HasADX = isSet(ebx7, cpuid_ADX)
   157  	X86.HasSHA = isSet(ebx7, cpuid_SHA)
   158  
   159  	X86.HasAVX512F = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
   160  	if X86.HasAVX512F {
   161  		X86.HasAVX512CD = isSet(ebx7, cpuid_AVX512CD)
   162  		X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
   163  		X86.HasAVX512DQ = isSet(ebx7, cpuid_AVX512DQ)
   164  		X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
   165  		X86.HasAVX512VPCLMULQDQ = isSet(ecx7, cpuid_AVX512VPCLMULQDQ)
   166  	}
   167  
   168  	X86.HasFSRM = isSet(edx7, cpuid_FSRM)
   169  
   170  	var maxExtendedInformation uint32
   171  	maxExtendedInformation, _, _, _ = cpuid(0x80000000, 0)
   172  
   173  	if maxExtendedInformation < 0x80000001 {
   174  		return
   175  	}
   176  
   177  	_, _, _, edxExt1 := cpuid(0x80000001, 0)
   178  	X86.HasRDTSCP = isSet(edxExt1, cpuid_RDTSCP)
   179  
   180  	doDerived = func() {
   181  		// Rather than carefully gating on fundamental AVX-512 features, we have
   182  		// a virtual "AVX512" feature that captures F+CD+BW+DQ+VL. BW, DQ, and
   183  		// VL have a huge effect on which AVX-512 instructions are available,
   184  		// and these have all been supported on everything except the earliest
   185  		// Phi chips with AVX-512. No CPU has had CD without F, so we include
   186  		// it. GOAMD64=v4 also implies exactly this set, and these are all
   187  		// included in AVX10.1.
   188  		X86.HasAVX512 = X86.HasAVX512F && X86.HasAVX512CD && X86.HasAVX512BW && X86.HasAVX512DQ && X86.HasAVX512VL
   189  	}
   190  }
   191  
   192  func isSet(hwc uint32, value uint32) bool {
   193  	return hwc&value != 0
   194  }
   195  
   196  // Name returns the CPU name given by the vendor.
   197  // If the CPU name can not be determined an
   198  // empty string is returned.
   199  func Name() string {
   200  	if maxExtendedFunctionInformation < 0x80000004 {
   201  		return ""
   202  	}
   203  
   204  	data := make([]byte, 0, 3*4*4)
   205  
   206  	var eax, ebx, ecx, edx uint32
   207  	eax, ebx, ecx, edx = cpuid(0x80000002, 0)
   208  	data = appendBytes(data, eax, ebx, ecx, edx)
   209  	eax, ebx, ecx, edx = cpuid(0x80000003, 0)
   210  	data = appendBytes(data, eax, ebx, ecx, edx)
   211  	eax, ebx, ecx, edx = cpuid(0x80000004, 0)
   212  	data = appendBytes(data, eax, ebx, ecx, edx)
   213  
   214  	// Trim leading spaces.
   215  	for len(data) > 0 && data[0] == ' ' {
   216  		data = data[1:]
   217  	}
   218  
   219  	// Trim tail after and including the first null byte.
   220  	for i, c := range data {
   221  		if c == '\x00' {
   222  			data = data[:i]
   223  			break
   224  		}
   225  	}
   226  
   227  	return string(data)
   228  }
   229  
   230  func appendBytes(b []byte, args ...uint32) []byte {
   231  	for _, arg := range args {
   232  		b = append(b,
   233  			byte((arg >> 0)),
   234  			byte((arg >> 8)),
   235  			byte((arg >> 16)),
   236  			byte((arg >> 24)))
   237  	}
   238  	return b
   239  }
   240
View as plain text