cpu: new package

This CL introduces a new cpu package for CPU/platform feature detection.
The cpu package is basically a copy of `internal/cpu` of the standard library.
Revision: bf86aec25972f3a100c3aa58a6abcbcc35bdea49

This CL does not export ARM64 and PPC64 feature detection since at the moment
ARM64/PPC64 requires standard library support.

Updates golang/go#24843

Change-Id: I11bc1ca60b116e902c941b5887c00870dbb1f899
Reviewed-on: https://go-review.googlesource.com/107015Reviewed-by: Brad Fitzpatrick <bradfitz@golang.org>

cpu/cpu.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// Package cpu implements processor feature detection for
+// various CPU architectures.
+package cpu
+
+// X86 contains the supported CPU features of the
+// current X86/AMD64 platform. If the current platform
+// is not X86/AMD64 then all feature flags are false.
+//
+// X86 is padded to avoid false sharing. Further the HasAVX
+// and HasAVX2 are only set if the OS supports XMM and YMM
+// registers in addition to the CPUID feature bit being set.
+var X86 struct {
+	_            [cacheLineSize]byte
+	HasAES       bool // AES hardware implementation (AES NI)
+	HasADX       bool // Multi-precision add-carry instruction extensions
+	HasAVX       bool // Advanced vector extension
+	HasAVX2      bool // Advanced vector extension 2
+	HasBMI1      bool // Bit manipulation instruction set 1
+	HasBMI2      bool // Bit manipulation instruction set 2
+	HasERMS      bool // Enhanced REP for MOVSB and STOSB
+	HasFMA       bool // Fused-multiply-add instructions
+	HasOSXSAVE   bool // OS supports XSAVE/XRESTOR for saving/restoring XMM registers.
+	HasPCLMULQDQ bool // PCLMULQDQ instruction - most often used for AES-GCM
+	HasPOPCNT    bool // Hamming weight instruction POPCNT.
+	HasSSE2      bool // Streaming SIMD extension 2 (always available on amd64)
+	HasSSE3      bool // Streaming SIMD extension 3
+	HasSSSE3     bool // Supplemental streaming SIMD extension 3
+	HasSSE41     bool // Streaming SIMD extension 4 and 4.1
+	HasSSE42     bool // Streaming SIMD extension 4 and 4.2
+	_            [cacheLineSize]byte
+}

cpu/cpu_arm.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cpu
+
+const cacheLineSize = 32

cpu/cpu_arm64.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cpu
+
+const cacheLineSize = 64

cpu/cpu_mips64x.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build mips64 mips64le
+
+package cpu
+
+const cacheLineSize = 32

cpu/cpu_mipsx.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build mips mipsle
+
+package cpu
+
+const cacheLineSize = 32

cpu/cpu_ppc64x.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build ppc64 ppc64le
+
+package cpu
+
+const cacheLineSize = 128

cpu/cpu_s390x.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cpu
+
+const cacheLineSize = 256

cpu/cpu_test.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+package cpu_test
+
+import (
+	"runtime"
+	"testing"
+
+	"golang.org/x/sys/cpu"
+)
+
+func TestAMD64minimalFeatures(t *testing.T) {
+	if runtime.GOARCH == "amd64" {
+		if !cpu.X86.HasSSE2 {
+			t.Fatalf("HasSSE2 expected true, got false")
+		}
+	}
+}
+
+func TestAVX2hasAVX(t *testing.T) {
+	if runtime.GOARCH == "amd64" {
+		if cpu.X86.HasAVX2 && !cpu.X86.HasAVX {
+			t.Fatalf("HasAVX expected true, got false")
+		}
+	}
+}

cpu/cpu_x86.go

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build 386 amd64 amd64p32
+
+package cpu
+
+const cacheLineSize = 64
+
+// cpuid is implemented in cpu_x86.s.
+func cpuid(eaxArg, ecxArg uint32) (eax, ebx, ecx, edx uint32)
+
+// xgetbv with ecx = 0 is implemented in cpu_x86.s.
+func xgetbv() (eax, edx uint32)
+
+func init() {
+	maxID, _, _, _ := cpuid(0, 0)
+
+	if maxID < 1 {
+		return
+	}
+
+	_, _, ecx1, edx1 := cpuid(1, 0)
+	X86.HasSSE2 = isSet(26, edx1)
+
+	X86.HasSSE3 = isSet(0, ecx1)
+	X86.HasPCLMULQDQ = isSet(1, ecx1)
+	X86.HasSSSE3 = isSet(9, ecx1)
+	X86.HasFMA = isSet(12, ecx1)
+	X86.HasSSE41 = isSet(19, ecx1)
+	X86.HasSSE42 = isSet(20, ecx1)
+	X86.HasPOPCNT = isSet(23, ecx1)
+	X86.HasAES = isSet(25, ecx1)
+	X86.HasOSXSAVE = isSet(27, ecx1)
+
+	osSupportsAVX := false
+	// For XGETBV, OSXSAVE bit is required and sufficient.
+	if X86.HasOSXSAVE {
+		eax, _ := xgetbv()
+		// Check if XMM and YMM registers have OS support.
+		osSupportsAVX = isSet(1, eax) && isSet(2, eax)
+	}
+
+	X86.HasAVX = isSet(28, ecx1) && osSupportsAVX
+
+	if maxID < 7 {
+		return
+	}
+
+	_, ebx7, _, _ := cpuid(7, 0)
+	X86.HasBMI1 = isSet(3, ebx7)
+	X86.HasAVX2 = isSet(5, ebx7) && osSupportsAVX
+	X86.HasBMI2 = isSet(8, ebx7)
+	X86.HasERMS = isSet(9, ebx7)
+	X86.HasADX = isSet(19, ebx7)
+}
+
+func isSet(bitpos uint, value uint32) bool {
+	return value&(1<<bitpos) != 0
+}

cpu/cpu_x86.s

+// Copyright 2018 The Go Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+// +build 386 amd64 amd64p32
+
+#include "textflag.h"
+
+// func cpuid(eaxArg, ecxArg uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·cpuid(SB), NOSPLIT, $0-24
+	MOVL eaxArg+0(FP), AX
+	MOVL ecxArg+4(FP), CX
+	CPUID
+	MOVL AX, eax+8(FP)
+	MOVL BX, ebx+12(FP)
+	MOVL CX, ecx+16(FP)
+	MOVL DX, edx+20(FP)
+	RET
+
+// func xgetbv() (eax, edx uint32)
+TEXT ·xgetbv(SB),NOSPLIT,$0-8
+	MOVL $0, CX
+	XGETBV
+	MOVL AX, eax+0(FP)
+	MOVL DX, edx+4(FP)
+	RET