diff --git a/cpu/cpu.go b/cpu/cpu.go
index abbec2d44b..c8df189daf 100644
--- a/cpu/cpu.go
+++ b/cpu/cpu.go
@@ -100,6 +100,7 @@ var ARM64 struct {
 	HasSHA512   bool // SHA512 hardware implementation
 	HasSVE      bool // Scalable Vector Extensions
 	HasASIMDFHM bool // Advanced SIMD multiplication FP16 to FP32
+	HasFMA      bool // Fused Multiply Add
 	_           CacheLinePad
 }
 
diff --git a/cpu/cpu_android_arm64.go b/cpu/cpu_android_arm64.go
new file mode 100644
index 0000000000..a6dbb15688
--- /dev/null
+++ b/cpu/cpu_android_arm64.go
@@ -0,0 +1,78 @@
+// Copyright 2021 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 arm64
+// +build android
+
+package cpu
+
+// HWCAP/HWCAP2 bits. These are exposed by Linux.
+const (
+	hwcap_FP       = 1 << 0
+	hwcap_ASIMD    = 1 << 1
+	hwcap_EVTSTRM  = 1 << 2
+	hwcap_AES      = 1 << 3
+	hwcap_PMULL    = 1 << 4
+	hwcap_SHA1     = 1 << 5
+	hwcap_SHA2     = 1 << 6
+	hwcap_CRC32    = 1 << 7
+	hwcap_ATOMICS  = 1 << 8
+	hwcap_FPHP     = 1 << 9
+	hwcap_ASIMDHP  = 1 << 10
+	hwcap_CPUID    = 1 << 11
+	hwcap_ASIMDRDM = 1 << 12
+	hwcap_JSCVT    = 1 << 13
+	hwcap_FCMA     = 1 << 14
+	hwcap_LRCPC    = 1 << 15
+	hwcap_DCPOP    = 1 << 16
+	hwcap_SHA3     = 1 << 17
+	hwcap_SM3      = 1 << 18
+	hwcap_SM4      = 1 << 19
+	hwcap_ASIMDDP  = 1 << 20
+	hwcap_SHA512   = 1 << 21
+	hwcap_SVE      = 1 << 22
+	hwcap_ASIMDFHM = 1 << 23
+)
+
+func osInit() {
+	if err := readHWCAP(); err != nil {
+		// failed to read /proc/self/auxv, try reading registers directly
+		readARM64Registers()
+		return
+	}
+
+	// Use HWCap information since reading aarch64 system registers
+	// is not supported in user space on older linux kernels.
+	ARM64.HasFP = isSet(hwCap, hwcap_FP)
+	ARM64.HasASIMD = isSet(hwCap, hwcap_ASIMD)
+	ARM64.HasEVTSTRM = isSet(hwCap, hwcap_EVTSTRM)
+	ARM64.HasAES = isSet(hwCap, hwcap_AES)
+	ARM64.HasPMULL = isSet(hwCap, hwcap_PMULL)
+	ARM64.HasSHA1 = isSet(hwCap, hwcap_SHA1)
+	ARM64.HasSHA2 = isSet(hwCap, hwcap_SHA2)
+	ARM64.HasCRC32 = isSet(hwCap, hwcap_CRC32)
+	ARM64.HasFPHP = isSet(hwCap, hwcap_FPHP)
+	ARM64.HasASIMDHP = isSet(hwCap, hwcap_ASIMDHP)
+	ARM64.HasASIMDRDM = isSet(hwCap, hwcap_ASIMDRDM)
+	ARM64.HasJSCVT = isSet(hwCap, hwcap_JSCVT)
+	ARM64.HasFCMA = isSet(hwCap, hwcap_FCMA)
+	ARM64.HasLRCPC = isSet(hwCap, hwcap_LRCPC)
+	ARM64.HasDCPOP = isSet(hwCap, hwcap_DCPOP)
+	ARM64.HasSHA3 = isSet(hwCap, hwcap_SHA3)
+	ARM64.HasSM3 = isSet(hwCap, hwcap_SM3)
+	ARM64.HasSM4 = isSet(hwCap, hwcap_SM4)
+	ARM64.HasASIMDDP = isSet(hwCap, hwcap_ASIMDDP)
+	ARM64.HasSHA512 = isSet(hwCap, hwcap_SHA512)
+	ARM64.HasSVE = isSet(hwCap, hwcap_SVE)
+	ARM64.HasASIMDFHM = isSet(hwCap, hwcap_ASIMDFHM)
+
+	// The Samsung S9+ kernel reports support for atomics, but not all cores
+	// actually support them, resulting in SIGILL. See issue #28431.
+	// TODO(elias.naur): Only disable the optimization on bad chipsets on android.
+	ARM64.HasATOMICS = false
+}
+
+func isSet(hwc uint, value uint) bool {
+	return hwc&value != 0
+}
diff --git a/cpu/cpu_arm64.go b/cpu/cpu_arm64.go
index 87dd5e3021..c9d9dd03e4 100644
--- a/cpu/cpu_arm64.go
+++ b/cpu/cpu_arm64.go
@@ -41,8 +41,8 @@ func archInit() {
 	switch runtime.GOOS {
 	case "freebsd":
 		readARM64Registers()
-	case "linux", "netbsd":
-		doinit()
+	case "linux", "netbsd", "android", "darwin":
+		osInit()
 	default:
 		// Most platforms don't seem to allow reading these registers.
 		//
@@ -52,13 +52,6 @@ func archInit() {
 	}
 }
 
-// setMinimalFeatures fakes the minimal ARM64 features expected by
-// TestARM64minimalFeatures.
-func setMinimalFeatures() {
-	ARM64.HasASIMD = true
-	ARM64.HasFP = true
-}
-
 func readARM64Registers() {
 	Initialized = true
 
@@ -170,3 +163,10 @@ func parseARM64SystemRegisters(isar0, isar1, pfr0 uint64) {
 func extractBits(data uint64, start, end uint) uint {
 	return (uint)(data>>start) & ((1 << (end - start + 1)) - 1)
 }
+
+// setMinimalFeatures fakes the minimal ARM64 features expected by
+// TestARM64minimalFeatures.
+func setMinimalFeatures() {
+	ARM64.HasASIMD = true
+	ARM64.HasFP = true
+}
diff --git a/cpu/cpu_darwin_arm64.go b/cpu/cpu_darwin_arm64.go
new file mode 100644
index 0000000000..be9124d9bf
--- /dev/null
+++ b/cpu/cpu_darwin_arm64.go
@@ -0,0 +1,74 @@
+// Copyright 2021 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 arm64
+// +build darwin
+// +build !ios
+
+package cpu
+
+import (
+	"unsafe"
+)
+
+const (
+	commpageHasNeonFP16      uint64 = 0x00000008         // ARM v8.2 NEON FP16 supported
+	commpageHasNeon          uint64 = 0x00000100         // Advanced SIMD is supported
+	commpageHasNeonHPFP      uint64 = 0x00000200         // Advanced SIMD half-precision
+	commpageHasVfp           uint64 = 0x00000400         // VFP is supported
+	commpageHasEvent         uint64 = 0x00001000         // WFE/SVE and period event wakeup
+	commpageHasFMA           uint64 = 0x00002000         // Fused multiply add is supported
+	commpageHasARMv82FHM     uint64 = 0x00004000         // Optional ARMv8.2 FMLAL/FMLSL instructions (required in ARMv8.4)
+	commpageHasARMv8Crypto   uint64 = 0x01000000         // Optional ARMv8 Crypto extensions
+	commpageHasARMv81Atomics uint64 = 0x02000000         // ARMv8.1 Atomic instructions supported
+	commpageHasARMv8Crc32    uint64 = 0x04000000         // Optional ARMv8 crc32 instructions (required in ARMv8.1)
+	commpageHasARMv82SHA512  uint64 = 0x80000000         // Optional ARMv8.2 SHA512 instructions
+	commpageHasARMv82SHA3    uint64 = 0x0000000100000000 // Optional ARMv8.2 SHA3 instructions
+)
+
+func osInit() {
+	ARM64.HasFP = darwinCheckFeatureEnabled(commpageHasVfp)
+	ARM64.HasASIMD = darwinCheckFeatureEnabled(commpageHasNeon)
+	ARM64.HasCRC32 = darwinCheckFeatureEnabled(commpageHasARMv8Crc32)
+	ARM64.HasATOMICS = darwinCheckFeatureEnabled(commpageHasARMv81Atomics)
+	ARM64.HasFPHP = darwinCheckFeatureEnabled(commpageHasNeonFP16)
+	ARM64.HasASIMDHP = darwinCheckFeatureEnabled(commpageHasNeonHPFP)
+	ARM64.HasSHA3 = darwinCheckFeatureEnabled(commpageHasARMv82SHA3)
+	ARM64.HasSHA512 = darwinCheckFeatureEnabled(commpageHasARMv82SHA512)
+	ARM64.HasASIMDFHM = darwinCheckFeatureEnabled(commpageHasARMv82FHM)
+	ARM64.HasSVE = darwinCheckFeatureEnabled(commpageHasEvent)
+	ARM64.HasFMA = darwinCheckFeatureEnabled(commpageHasFMA)
+
+	// There are no hw.optional sysctl values for the below features on Mac OS 11.0
+	// to detect their supported state dynamically. Assume the CPU features that
+	// Apple Silicon M1 supports to be available as a minimal set of features
+	// to all Go programs running on darwin/arm64.
+	ARM64.HasEVTSTRM = true
+	ARM64.HasAES = true
+	ARM64.HasPMULL = true
+	ARM64.HasSHA1 = true
+	ARM64.HasSHA2 = true
+	ARM64.HasCPUID = true
+	ARM64.HasASIMDRDM = true
+	ARM64.HasJSCVT = true
+	ARM64.HasFCMA = true
+	ARM64.HasLRCPC = true
+	ARM64.HasDCPOP = true
+	ARM64.HasSM3 = true
+	ARM64.HasSM4 = true
+	ARM64.HasASIMDDP = true
+}
+
+const (
+	commpage64_base_address       = 0x0000000fffffc000
+	commpage64_cpu_capabilities64 = (commpage64_base_address + 0x010)
+)
+
+func darwinCheckFeatureEnabled(feature_vec uint64) bool {
+	if (*(*int64)(unsafe.Pointer(uintptr(commpage64_cpu_capabilities64))))&int64(feature_vec) > 0 {
+		return true
+	}
+
+	return false
+}
diff --git a/cpu/cpu_linux_arm64.go b/cpu/cpu_linux_arm64.go
index 79a38a0b9b..7f55a6ed89 100644
--- a/cpu/cpu_linux_arm64.go
+++ b/cpu/cpu_linux_arm64.go
@@ -2,6 +2,10 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
+// +build arm64
+// +build linux
+// +build !android
+
 package cpu
 
 // HWCAP/HWCAP2 bits. These are exposed by Linux.
@@ -32,7 +36,7 @@ const (
 	hwcap_ASIMDFHM = 1 << 23
 )
 
-func doinit() {
+func osInit() {
 	if err := readHWCAP(); err != nil {
 		// failed to read /proc/self/auxv, try reading registers directly
 		readARM64Registers()
diff --git a/cpu/cpu_other_arm64.go b/cpu/cpu_other_arm64.go
index f8c484f589..e5a23d1ac3 100644
--- a/cpu/cpu_other_arm64.go
+++ b/cpu/cpu_other_arm64.go
@@ -2,9 +2,11 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-//go:build !linux && !netbsd && arm64
-// +build !linux,!netbsd,arm64
+//go:build !linux && !netbsd && !darwin && arm64
+// +build !linux,!netbsd,!darwin,arm64
 
 package cpu
 
-func doinit() {}
+func osInit() {
+	setMinimalFeatures()
+}
diff --git a/cpu/cpu_test.go b/cpu/cpu_test.go
index 5f7f843fe4..ba25551509 100644
--- a/cpu/cpu_test.go
+++ b/cpu/cpu_test.go
@@ -42,7 +42,7 @@ func TestAVX512HasAVX2AndAVX(t *testing.T) {
 }
 
 func TestARM64minimalFeatures(t *testing.T) {
-	if runtime.GOARCH != "arm64" || (runtime.GOOS == "darwin" || runtime.GOOS == "ios") {
+	if runtime.GOARCH != "arm64" || runtime.GOOS == "ios" {
 		return
 	}
 	if !cpu.ARM64.HasASIMD {