Merge branch 'master' into windows-conin

Author: aymanbagabas

README.md

@@ -5,16 +5,13 @@
 This repository holds supplemental Go packages for low-level interactions with
 the operating system.
 
-## Download/Install
-
-The easiest way to install is to run `go get -u golang.org/x/sys`. You can
-also manually git clone the repository to `$GOPATH/src/golang.org/x/sys`.
-
 ## Report Issues / Send Patches
 
 This repository uses Gerrit for code changes. To learn how to submit changes to
-this repository, see https://golang.org/doc/contribute.html.
+this repository, see https://go.dev/doc/contribute.
+
+The git repository is https://go.googlesource.com/sys.
 
 The main issue tracker for the sys repository is located at
-https://github.com/golang/go/issues. Prefix your issue with "x/sys:" in the
+https://go.dev/issues. Prefix your issue with "x/sys:" in the
 subject line, so it is easy to find.

cpu/asm_darwin_x86_gc.s

@@ -0,0 +1,17 @@
+// Copyright 2024 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.
+
+//go:build darwin && amd64 && gc
+
+#include "textflag.h"
+
+TEXT libc_sysctl_trampoline<>(SB),NOSPLIT,$0-0
+	JMP	libc_sysctl(SB)
+GLOBL	·libc_sysctl_trampoline_addr(SB), RODATA, $8
+DATA	·libc_sysctl_trampoline_addr(SB)/8, $libc_sysctl_trampoline<>(SB)
+
+TEXT libc_sysctlbyname_trampoline<>(SB),NOSPLIT,$0-0
+	JMP	libc_sysctlbyname(SB)
+GLOBL	·libc_sysctlbyname_trampoline_addr(SB), RODATA, $8
+DATA	·libc_sysctlbyname_trampoline_addr(SB)/8, $libc_sysctlbyname_trampoline<>(SB)

cpu/cpu.go

@@ -72,6 +72,9 @@ var X86 struct {
 	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
+	HasAVXIFMA          bool // Advanced vector extension Integer Fused Multiply Add
+	HasAVXVNNI          bool // Advanced vector extension Vector Neural Network Instructions
+	HasAVXVNNIInt8      bool // Advanced vector extension Vector Neural Network Int8 instructions
 	_                   CacheLinePad
 }
 

cpu/cpu_darwin_x86.go

@@ -0,0 +1,61 @@
+// Copyright 2024 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.
+
+//go:build darwin && amd64 && gc
+
+package cpu
+
+// darwinSupportsAVX512 checks Darwin kernel for AVX512 support via sysctl
+// call (see issue 43089). It also restricts AVX512 support for Darwin to
+// kernel version 21.3.0 (MacOS 12.2.0) or later (see issue 49233).
+//
+// Background:
+// Darwin implements a special mechanism to economize on thread state when
+// AVX512 specific registers are not in use. This scheme minimizes state when
+// preempting threads that haven't yet used any AVX512 instructions, but adds
+// special requirements to check for AVX512 hardware support at runtime (e.g.
+// via sysctl call or commpage inspection). See issue 43089 and link below for
+// full background:
+// https://github.com/apple-oss-distributions/xnu/blob/xnu-11215.1.10/osfmk/i386/fpu.c#L214-L240
+//
+// Additionally, all versions of the Darwin kernel from 19.6.0 through 21.2.0
+// (corresponding to MacOS 10.15.6 - 12.1) have a bug that can cause corruption
+// of the AVX512 mask registers (K0-K7) upon signal return. For this reason
+// AVX512 is considered unsafe to use on Darwin for kernel versions prior to
+// 21.3.0, where a fix has been confirmed. See issue 49233 for full background.
+func darwinSupportsAVX512() bool {
+	return darwinSysctlEnabled([]byte("hw.optional.avx512f\x00")) && darwinKernelVersionCheck(21, 3, 0)
+}
+
+// Ensure Darwin kernel version is at least major.minor.patch, avoiding dependencies
+func darwinKernelVersionCheck(major, minor, patch int) bool {
+	var release [256]byte
+	err := darwinOSRelease(&release)
+	if err != nil {
+		return false
+	}
+
+	var mmp [3]int
+	c := 0
+Loop:
+	for _, b := range release[:] {
+		switch {
+		case b >= '0' && b <= '9':
+			mmp[c] = 10*mmp[c] + int(b-'0')
+		case b == '.':
+			c++
+			if c > 2 {
+				return false
+			}
+		case b == 0:
+			break Loop
+		default:
+			return false
+		}
+	}
+	if c != 2 {
+		return false
+	}
+	return mmp[0] > major || mmp[0] == major && (mmp[1] > minor || mmp[1] == minor && mmp[2] >= patch)
+}

cpu/cpu_gc_x86.go

@@ -6,10 +6,10 @@
 
 package cpu
 
-// cpuid is implemented in cpu_x86.s for gc compiler
+// cpuid is implemented in cpu_gc_x86.s for gc compiler
 // and in cpu_gccgo.c for gccgo.
 func cpuid(eaxArg, ecxArg uint32) (eax, ebx, ecx, edx uint32)
 
-// xgetbv with ecx = 0 is implemented in cpu_x86.s for gc compiler
+// xgetbv with ecx = 0 is implemented in cpu_gc_x86.s for gc compiler
 // and in cpu_gccgo.c for gccgo.
 func xgetbv() (eax, edx uint32)

cpu/cpu_x86.s renamed to cpu/cpu_gc_x86.s

@@ -18,7 +18,7 @@ TEXT ·cpuid(SB), NOSPLIT, $0-24
 	RET
 
 // func xgetbv() (eax, edx uint32)
-TEXT ·xgetbv(SB),NOSPLIT,$0-8
+TEXT ·xgetbv(SB), NOSPLIT, $0-8
 	MOVL $0, CX
 	XGETBV
 	MOVL AX, eax+0(FP)

cpu/cpu_gccgo_x86.go

@@ -23,9 +23,3 @@ func xgetbv() (eax, edx uint32) {
 	gccgoXgetbv(&a, &d)
 	return a, d
 }
-
-// gccgo doesn't build on Darwin, per:
-// https://github.com/Homebrew/homebrew-core/blob/HEAD/Formula/gcc.rb#L76
-func darwinSupportsAVX512() bool {
-	return false
-}

cpu/cpu_other_x86.go

@@ -0,0 +1,11 @@
+// Copyright 2024 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.
+
+//go:build 386 || amd64p32 || (amd64 && (!darwin || !gc))
+
+package cpu
+
+func darwinSupportsAVX512() bool {
+	panic("only implemented for gc && amd64 && darwin")
+}

cpu/cpu_test.go

@@ -41,6 +41,40 @@ func TestAVX512HasAVX2AndAVX(t *testing.T) {
 	}
 }
 
+func TestAVX512BF16HasAVX512(t *testing.T) {
+	if runtime.GOARCH == "amd64" {
+		if cpu.X86.HasAVX512BF16 && !cpu.X86.HasAVX512 {
+			t.Fatal("HasAVX512 expected true, got false")
+		}
+	}
+}
+
+func TestAVXVNNIHasAVX(t *testing.T) {
+	if cpu.X86.HasAVXVNNI && !cpu.X86.HasAVX {
+		t.Fatal("HasAVX expected true, got false")
+	}
+}
+
+func TestAVXIFMAHasAVXVNNIAndAVX(t *testing.T) {
+	if cpu.X86.HasAVXIFMA && !cpu.X86.HasAVX {
+		t.Fatal("HasAVX expected true, got false")
+	}
+
+	if cpu.X86.HasAVXIFMA && !cpu.X86.HasAVXVNNI {
+		t.Fatal("HasAVXVNNI expected true, got false")
+	}
+}
+
+func TestAVXVNNIInt8HasAVXVNNIAndAVX(t *testing.T) {
+	if cpu.X86.HasAVXVNNIInt8 && !cpu.X86.HasAVXVNNI {
+		t.Fatal("HasAVXVNNI expected true, got false")
+	}
+
+	if cpu.X86.HasAVXVNNIInt8 && !cpu.X86.HasAVX {
+		t.Fatal("HasAVX expected true, got false")
+	}
+}
+
 func TestARM64minimalFeatures(t *testing.T) {
 	if runtime.GOARCH != "arm64" || runtime.GOOS == "ios" {
 		return

cpu/cpu_x86.go

@@ -53,6 +53,9 @@ func initOptions() {
 		{Name: "sse41", Feature: &X86.HasSSE41},
 		{Name: "sse42", Feature: &X86.HasSSE42},
 		{Name: "ssse3", Feature: &X86.HasSSSE3},
+		{Name: "avxifma", Feature: &X86.HasAVXIFMA},
+		{Name: "avxvnni", Feature: &X86.HasAVXVNNI},
+		{Name: "avxvnniint8", Feature: &X86.HasAVXVNNIInt8},
 
 		// These capabilities should always be enabled on amd64:
 		{Name: "sse2", Feature: &X86.HasSSE2, Required: runtime.GOARCH == "amd64"},
@@ -92,10 +95,8 @@ func archInit() {
 		osSupportsAVX = isSet(1, eax) && isSet(2, eax)
 
 		if runtime.GOOS == "darwin" {
-			// Darwin doesn't save/restore AVX-512 mask registers correctly across signal handlers.
-			// Since users can't rely on mask register contents, let's not advertise AVX-512 support.
-			// See issue 49233.
-			osSupportsAVX512 = false
+			// Darwin requires special AVX512 checks, see cpu_darwin_x86.go
+			osSupportsAVX512 = osSupportsAVX && darwinSupportsAVX512()
 		} else {
 			// Check if OPMASK and ZMM registers have OS support.
 			osSupportsAVX512 = osSupportsAVX && isSet(5, eax) && isSet(6, eax) && isSet(7, eax)
@@ -108,7 +109,7 @@ func archInit() {
 		return
 	}
 
-	_, ebx7, ecx7, edx7 := cpuid(7, 0)
+	eax7, ebx7, ecx7, edx7 := cpuid(7, 0)
 	X86.HasBMI1 = isSet(3, ebx7)
 	X86.HasAVX2 = isSet(5, ebx7) && osSupportsAVX
 	X86.HasBMI2 = isSet(8, ebx7)
@@ -136,14 +137,24 @@ func archInit() {
 		X86.HasAVX512VAES = isSet(9, ecx7)
 		X86.HasAVX512VBMI2 = isSet(6, ecx7)
 		X86.HasAVX512BITALG = isSet(12, ecx7)
-
-		eax71, _, _, _ := cpuid(7, 1)
-		X86.HasAVX512BF16 = isSet(5, eax71)
 	}
 
 	X86.HasAMXTile = isSet(24, edx7)
 	X86.HasAMXInt8 = isSet(25, edx7)
 	X86.HasAMXBF16 = isSet(22, edx7)
+
+	// These features depend on the second level of extended features.
+	if eax7 >= 1 {
+		eax71, _, _, edx71 := cpuid(7, 1)
+		if X86.HasAVX512 {
+			X86.HasAVX512BF16 = isSet(5, eax71)
+		}
+		if X86.HasAVX {
+			X86.HasAVXIFMA = isSet(23, eax71)
+			X86.HasAVXVNNI = isSet(4, eax71)
+			X86.HasAVXVNNIInt8 = isSet(4, edx71)
+		}
+	}
 }
 
 func isSet(bitpos uint, value uint32) bool {