Adding upstream version 1.19.8.upstream/1.19.8upstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

51 files changed, 6724 insertions, 0 deletions

diff --git a/src/runtime/internal/atomic/atomic_386.go b/src/runtime/internal/atomic/atomic_386.go
new file mode 100644
index 0000000..bf2f4b9
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_386.go
@@ -0,0 +1,103 @@
+// Copyright 2009 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
+
+package atomic
+
+import "unsafe"
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Load
+//go:linkname Loadp
+
+//go:nosplit
+//go:noinline
+func Load(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer {
+	return *(*unsafe.Pointer)(ptr)
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcquintptr(ptr *uintptr) uintptr {
+	return *ptr
+}
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+//go:nosplit
+//go:noinline
+func Load8(ptr *uint8) uint8 {
+	return *ptr
+}
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
diff --git a/src/runtime/internal/atomic/atomic_386.s b/src/runtime/internal/atomic/atomic_386.s
new file mode 100644
index 0000000..724d515
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_386.s
@@ -0,0 +1,285 @@
+// Copyright 2015 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.
+
+#include "textflag.h"
+#include "funcdata.h"
+
+// bool Cas(int32 *val, int32 old, int32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	}else
+//		return 0;
+TEXT ·Cas(SB), NOSPLIT, $0-13
+	MOVL	ptr+0(FP), BX
+	MOVL	old+4(FP), AX
+	MOVL	new+8(FP), CX
+	LOCK
+	CMPXCHGL	CX, 0(BX)
+	SETEQ	ret+12(FP)
+	RET
+
+TEXT ·Casint32(SB), NOSPLIT, $0-13
+	JMP	·Cas(SB)
+
+TEXT ·Casint64(SB), NOSPLIT, $0-21
+	JMP	·Cas64(SB)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-13
+	JMP	·Cas(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-13
+	JMP	·Cas(SB)
+
+TEXT ·Loaduintptr(SB), NOSPLIT, $0-8
+	JMP	·Load(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT, $0-8
+	JMP	·Load(SB)
+
+TEXT ·Storeint32(SB), NOSPLIT, $0-8
+	JMP	·Store(SB)
+
+TEXT ·Storeint64(SB), NOSPLIT, $0-12
+	JMP	·Store64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-8
+	JMP	·Store(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-12
+	JMP	·Xadd(SB)
+
+TEXT ·Loadint32(SB), NOSPLIT, $0-8
+	JMP	·Load(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-12
+	JMP	·Load64(SB)
+
+TEXT ·Xaddint32(SB), NOSPLIT, $0-12
+	JMP	·Xadd(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-20
+	JMP	·Xadd64(SB)
+
+// bool ·Cas64(uint64 *val, uint64 old, uint64 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		return 0;
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-21
+	NO_LOCAL_POINTERS
+	MOVL	ptr+0(FP), BP
+	TESTL	$7, BP
+	JZ	2(PC)
+	CALL	·panicUnaligned(SB)
+	MOVL	old_lo+4(FP), AX
+	MOVL	old_hi+8(FP), DX
+	MOVL	new_lo+12(FP), BX
+	MOVL	new_hi+16(FP), CX
+	LOCK
+	CMPXCHG8B	0(BP)
+	SETEQ	ret+20(FP)
+	RET
+
+// bool Casp1(void **p, void *old, void *new)
+// Atomically:
+//	if(*p == old){
+//		*p = new;
+//		return 1;
+//	}else
+//		return 0;
+TEXT ·Casp1(SB), NOSPLIT, $0-13
+	MOVL	ptr+0(FP), BX
+	MOVL	old+4(FP), AX
+	MOVL	new+8(FP), CX
+	LOCK
+	CMPXCHGL	CX, 0(BX)
+	SETEQ	ret+12(FP)
+	RET
+
+// uint32 Xadd(uint32 volatile *val, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-12
+	MOVL	ptr+0(FP), BX
+	MOVL	delta+4(FP), AX
+	MOVL	AX, CX
+	LOCK
+	XADDL	AX, 0(BX)
+	ADDL	CX, AX
+	MOVL	AX, ret+8(FP)
+	RET
+
+TEXT ·Xadd64(SB), NOSPLIT, $0-20
+	NO_LOCAL_POINTERS
+	// no XADDQ so use CMPXCHG8B loop
+	MOVL	ptr+0(FP), BP
+	TESTL	$7, BP
+	JZ	2(PC)
+	CALL	·panicUnaligned(SB)
+	// DI:SI = delta
+	MOVL	delta_lo+4(FP), SI
+	MOVL	delta_hi+8(FP), DI
+	// DX:AX = *addr
+	MOVL	0(BP), AX
+	MOVL	4(BP), DX
+addloop:
+	// CX:BX = DX:AX (*addr) + DI:SI (delta)
+	MOVL	AX, BX
+	MOVL	DX, CX
+	ADDL	SI, BX
+	ADCL	DI, CX
+
+	// if *addr == DX:AX {
+	//	*addr = CX:BX
+	// } else {
+	//	DX:AX = *addr
+	// }
+	// all in one instruction
+	LOCK
+	CMPXCHG8B	0(BP)
+
+	JNZ	addloop
+
+	// success
+	// return CX:BX
+	MOVL	BX, ret_lo+12(FP)
+	MOVL	CX, ret_hi+16(FP)
+	RET
+
+TEXT ·Xchg(SB), NOSPLIT, $0-12
+	MOVL	ptr+0(FP), BX
+	MOVL	new+4(FP), AX
+	XCHGL	AX, 0(BX)
+	MOVL	AX, ret+8(FP)
+	RET
+
+TEXT ·Xchgint32(SB), NOSPLIT, $0-12
+	JMP	·Xchg(SB)
+
+TEXT ·Xchgint64(SB), NOSPLIT, $0-20
+	JMP	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-12
+	JMP	·Xchg(SB)
+
+TEXT ·Xchg64(SB),NOSPLIT,$0-20
+	NO_LOCAL_POINTERS
+	// no XCHGQ so use CMPXCHG8B loop
+	MOVL	ptr+0(FP), BP
+	TESTL	$7, BP
+	JZ	2(PC)
+	CALL	·panicUnaligned(SB)
+	// CX:BX = new
+	MOVL	new_lo+4(FP), BX
+	MOVL	new_hi+8(FP), CX
+	// DX:AX = *addr
+	MOVL	0(BP), AX
+	MOVL	4(BP), DX
+swaploop:
+	// if *addr == DX:AX
+	//	*addr = CX:BX
+	// else
+	//	DX:AX = *addr
+	// all in one instruction
+	LOCK
+	CMPXCHG8B	0(BP)
+	JNZ	swaploop
+
+	// success
+	// return DX:AX
+	MOVL	AX, ret_lo+12(FP)
+	MOVL	DX, ret_hi+16(FP)
+	RET
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-8
+	MOVL	ptr+0(FP), BX
+	MOVL	val+4(FP), AX
+	XCHGL	AX, 0(BX)
+	RET
+
+TEXT ·Store(SB), NOSPLIT, $0-8
+	MOVL	ptr+0(FP), BX
+	MOVL	val+4(FP), AX
+	XCHGL	AX, 0(BX)
+	RET
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-8
+	JMP	·Store(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-8
+	JMP	·Store(SB)
+
+// uint64 atomicload64(uint64 volatile* addr);
+TEXT ·Load64(SB), NOSPLIT, $0-12
+	NO_LOCAL_POINTERS
+	MOVL	ptr+0(FP), AX
+	TESTL	$7, AX
+	JZ	2(PC)
+	CALL	·panicUnaligned(SB)
+	MOVQ	(AX), M0
+	MOVQ	M0, ret+4(FP)
+	EMMS
+	RET
+
+// void ·Store64(uint64 volatile* addr, uint64 v);
+TEXT ·Store64(SB), NOSPLIT, $0-12
+	NO_LOCAL_POINTERS
+	MOVL	ptr+0(FP), AX
+	TESTL	$7, AX
+	JZ	2(PC)
+	CALL	·panicUnaligned(SB)
+	// MOVQ and EMMS were introduced on the Pentium MMX.
+	MOVQ	val+4(FP), M0
+	MOVQ	M0, (AX)
+	EMMS
+	// This is essentially a no-op, but it provides required memory fencing.
+	// It can be replaced with MFENCE, but MFENCE was introduced only on the Pentium4 (SSE2).
+	XORL	AX, AX
+	LOCK
+	XADDL	AX, (SP)
+	RET
+
+// void	·Or8(byte volatile*, byte);
+TEXT ·Or8(SB), NOSPLIT, $0-5
+	MOVL	ptr+0(FP), AX
+	MOVB	val+4(FP), BX
+	LOCK
+	ORB	BX, (AX)
+	RET
+
+// void	·And8(byte volatile*, byte);
+TEXT ·And8(SB), NOSPLIT, $0-5
+	MOVL	ptr+0(FP), AX
+	MOVB	val+4(FP), BX
+	LOCK
+	ANDB	BX, (AX)
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-5
+	MOVL	ptr+0(FP), BX
+	MOVB	val+4(FP), AX
+	XCHGB	AX, 0(BX)
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-8
+	MOVL	ptr+0(FP), AX
+	MOVL	val+4(FP), BX
+	LOCK
+	ORL	BX, (AX)
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-8
+	MOVL	ptr+0(FP), AX
+	MOVL	val+4(FP), BX
+	LOCK
+	ANDL	BX, (AX)
+	RET
diff --git a/src/runtime/internal/atomic/atomic_amd64.go b/src/runtime/internal/atomic/atomic_amd64.go
new file mode 100644
index 0000000..52a8362
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_amd64.go
@@ -0,0 +1,117 @@
+// Copyright 2009 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 atomic
+
+import "unsafe"
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Load
+//go:linkname Loadp
+//go:linkname Load64
+
+//go:nosplit
+//go:noinline
+func Load(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer {
+	return *(*unsafe.Pointer)(ptr)
+}
+
+//go:nosplit
+//go:noinline
+func Load64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcquintptr(ptr *uintptr) uintptr {
+	return *ptr
+}
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:nosplit
+//go:noinline
+func Load8(ptr *uint8) uint8 {
+	return *ptr
+}
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreRel64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
+
+// StorepNoWB performs *ptr = val atomically and without a write
+// barrier.
+//
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
diff --git a/src/runtime/internal/atomic/atomic_amd64.s b/src/runtime/internal/atomic/atomic_amd64.s
new file mode 100644
index 0000000..d21514b
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_amd64.s
@@ -0,0 +1,225 @@
+// Copyright 2015 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.
+
+// Note: some of these functions are semantically inlined
+// by the compiler (in src/cmd/compile/internal/gc/ssa.go).
+
+#include "textflag.h"
+
+TEXT ·Loaduintptr(SB), NOSPLIT, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Loadint32(SB), NOSPLIT, $0-12
+	JMP	·Load(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	JMP	·Load64(SB)
+
+// bool Cas(int32 *val, int32 old, int32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB),NOSPLIT,$0-17
+	MOVQ	ptr+0(FP), BX
+	MOVL	old+8(FP), AX
+	MOVL	new+12(FP), CX
+	LOCK
+	CMPXCHGL	CX, 0(BX)
+	SETEQ	ret+16(FP)
+	RET
+
+// bool	·Cas64(uint64 *val, uint64 old, uint64 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		return 0;
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVQ	ptr+0(FP), BX
+	MOVQ	old+8(FP), AX
+	MOVQ	new+16(FP), CX
+	LOCK
+	CMPXCHGQ	CX, 0(BX)
+	SETEQ	ret+24(FP)
+	RET
+
+// bool Casp1(void **val, void *old, void *new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	MOVQ	ptr+0(FP), BX
+	MOVQ	old+8(FP), AX
+	MOVQ	new+16(FP), CX
+	LOCK
+	CMPXCHGQ	CX, 0(BX)
+	SETEQ	ret+24(FP)
+	RET
+
+TEXT ·Casint32(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+TEXT ·Casint64(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+// uint32 Xadd(uint32 volatile *val, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVQ	ptr+0(FP), BX
+	MOVL	delta+8(FP), AX
+	MOVL	AX, CX
+	LOCK
+	XADDL	AX, 0(BX)
+	ADDL	CX, AX
+	MOVL	AX, ret+16(FP)
+	RET
+
+// uint64 Xadd64(uint64 volatile *val, int64 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVQ	ptr+0(FP), BX
+	MOVQ	delta+8(FP), AX
+	MOVQ	AX, CX
+	LOCK
+	XADDQ	AX, 0(BX)
+	ADDQ	CX, AX
+	MOVQ	AX, ret+16(FP)
+	RET
+
+TEXT ·Xaddint32(SB), NOSPLIT, $0-20
+	JMP	·Xadd(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+// uint32 Xchg(ptr *uint32, new uint32)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVQ	ptr+0(FP), BX
+	MOVL	new+8(FP), AX
+	XCHGL	AX, 0(BX)
+	MOVL	AX, ret+16(FP)
+	RET
+
+// uint64 Xchg64(ptr *uint64, new uint64)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVQ	ptr+0(FP), BX
+	MOVQ	new+8(FP), AX
+	XCHGQ	AX, 0(BX)
+	MOVQ	AX, ret+16(FP)
+	RET
+
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	JMP	·Xchg(SB)
+
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	MOVQ	ptr+0(FP), BX
+	MOVQ	val+8(FP), AX
+	XCHGQ	AX, 0(BX)
+	RET
+
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOVQ	ptr+0(FP), BX
+	MOVL	val+8(FP), AX
+	XCHGL	AX, 0(BX)
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOVQ	ptr+0(FP), BX
+	MOVB	val+8(FP), AX
+	XCHGB	AX, 0(BX)
+	RET
+
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOVQ	ptr+0(FP), BX
+	MOVQ	val+8(FP), AX
+	XCHGQ	AX, 0(BX)
+	RET
+
+TEXT ·Storeint32(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·Storeint64(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·StoreRel64(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+// void	·Or8(byte volatile*, byte);
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVQ	ptr+0(FP), AX
+	MOVB	val+8(FP), BX
+	LOCK
+	ORB	BX, (AX)
+	RET
+
+// void	·And8(byte volatile*, byte);
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVQ	ptr+0(FP), AX
+	MOVB	val+8(FP), BX
+	LOCK
+	ANDB	BX, (AX)
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVQ	ptr+0(FP), AX
+	MOVL	val+8(FP), BX
+	LOCK
+	ORL	BX, (AX)
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVQ	ptr+0(FP), AX
+	MOVL	val+8(FP), BX
+	LOCK
+	ANDL	BX, (AX)
+	RET
diff --git a/src/runtime/internal/atomic/atomic_arm.go b/src/runtime/internal/atomic/atomic_arm.go
new file mode 100644
index 0000000..bdb1847
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_arm.go
@@ -0,0 +1,244 @@
+// Copyright 2009 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 arm
+
+package atomic
+
+import (
+	"internal/cpu"
+	"unsafe"
+)
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Xchg
+//go:linkname Xchguintptr
+
+type spinlock struct {
+	v uint32
+}
+
+//go:nosplit
+func (l *spinlock) lock() {
+	for {
+		if Cas(&l.v, 0, 1) {
+			return
+		}
+	}
+}
+
+//go:nosplit
+func (l *spinlock) unlock() {
+	Store(&l.v, 0)
+}
+
+var locktab [57]struct {
+	l   spinlock
+	pad [cpu.CacheLinePadSize - unsafe.Sizeof(spinlock{})]byte
+}
+
+func addrLock(addr *uint64) *spinlock {
+	return &locktab[(uintptr(unsafe.Pointer(addr))>>3)%uintptr(len(locktab))].l
+}
+
+// Atomic add and return new value.
+//
+//go:nosplit
+func Xadd(val *uint32, delta int32) uint32 {
+	for {
+		oval := *val
+		nval := oval + uint32(delta)
+		if Cas(val, oval, nval) {
+			return nval
+		}
+	}
+}
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:nosplit
+func Xchg(addr *uint32, v uint32) uint32 {
+	for {
+		old := *addr
+		if Cas(addr, old, v) {
+			return old
+		}
+	}
+}
+
+//go:nosplit
+func Xchguintptr(addr *uintptr, v uintptr) uintptr {
+	return uintptr(Xchg((*uint32)(unsafe.Pointer(addr)), uint32(v)))
+}
+
+// Not noescape -- it installs a pointer to addr.
+func StorepNoWB(addr unsafe.Pointer, v unsafe.Pointer)
+
+//go:noescape
+func Store(addr *uint32, v uint32)
+
+//go:noescape
+func StoreRel(addr *uint32, v uint32)
+
+//go:noescape
+func StoreReluintptr(addr *uintptr, v uintptr)
+
+//go:nosplit
+func goCas64(addr *uint64, old, new uint64) bool {
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		*(*int)(nil) = 0 // crash on unaligned uint64
+	}
+	_ = *addr // if nil, fault before taking the lock
+	var ok bool
+	addrLock(addr).lock()
+	if *addr == old {
+		*addr = new
+		ok = true
+	}
+	addrLock(addr).unlock()
+	return ok
+}
+
+//go:nosplit
+func goXadd64(addr *uint64, delta int64) uint64 {
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		*(*int)(nil) = 0 // crash on unaligned uint64
+	}
+	_ = *addr // if nil, fault before taking the lock
+	var r uint64
+	addrLock(addr).lock()
+	r = *addr + uint64(delta)
+	*addr = r
+	addrLock(addr).unlock()
+	return r
+}
+
+//go:nosplit
+func goXchg64(addr *uint64, v uint64) uint64 {
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		*(*int)(nil) = 0 // crash on unaligned uint64
+	}
+	_ = *addr // if nil, fault before taking the lock
+	var r uint64
+	addrLock(addr).lock()
+	r = *addr
+	*addr = v
+	addrLock(addr).unlock()
+	return r
+}
+
+//go:nosplit
+func goLoad64(addr *uint64) uint64 {
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		*(*int)(nil) = 0 // crash on unaligned uint64
+	}
+	_ = *addr // if nil, fault before taking the lock
+	var r uint64
+	addrLock(addr).lock()
+	r = *addr
+	addrLock(addr).unlock()
+	return r
+}
+
+//go:nosplit
+func goStore64(addr *uint64, v uint64) {
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		*(*int)(nil) = 0 // crash on unaligned uint64
+	}
+	_ = *addr // if nil, fault before taking the lock
+	addrLock(addr).lock()
+	*addr = v
+	addrLock(addr).unlock()
+}
+
+//go:nosplit
+func Or8(addr *uint8, v uint8) {
+	// Align down to 4 bytes and use 32-bit CAS.
+	uaddr := uintptr(unsafe.Pointer(addr))
+	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))
+	word := uint32(v) << ((uaddr & 3) * 8) // little endian
+	for {
+		old := *addr32
+		if Cas(addr32, old, old|word) {
+			return
+		}
+	}
+}
+
+//go:nosplit
+func And8(addr *uint8, v uint8) {
+	// Align down to 4 bytes and use 32-bit CAS.
+	uaddr := uintptr(unsafe.Pointer(addr))
+	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))
+	word := uint32(v) << ((uaddr & 3) * 8)    // little endian
+	mask := uint32(0xFF) << ((uaddr & 3) * 8) // little endian
+	word |= ^mask
+	for {
+		old := *addr32
+		if Cas(addr32, old, old&word) {
+			return
+		}
+	}
+}
+
+//go:nosplit
+func Or(addr *uint32, v uint32) {
+	for {
+		old := *addr
+		if Cas(addr, old, old|v) {
+			return
+		}
+	}
+}
+
+//go:nosplit
+func And(addr *uint32, v uint32) {
+	for {
+		old := *addr
+		if Cas(addr, old, old&v) {
+			return
+		}
+	}
+}
+
+//go:nosplit
+func armcas(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Load(addr *uint32) uint32
+
+// NO go:noescape annotation; *addr escapes if result escapes (#31525)
+func Loadp(addr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func Load8(addr *uint8) uint8
+
+//go:noescape
+func LoadAcq(addr *uint32) uint32
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func Cas64(addr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(addr *uint32, old, new uint32) bool
+
+//go:noescape
+func Xadd64(addr *uint64, delta int64) uint64
+
+//go:noescape
+func Xchg64(addr *uint64, v uint64) uint64
+
+//go:noescape
+func Load64(addr *uint64) uint64
+
+//go:noescape
+func Store8(addr *uint8, v uint8)
+
+//go:noescape
+func Store64(addr *uint64, v uint64)
diff --git a/src/runtime/internal/atomic/atomic_arm.s b/src/runtime/internal/atomic/atomic_arm.s
new file mode 100644
index 0000000..92cbe8a
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_arm.s
@@ -0,0 +1,297 @@
+// Copyright 2015 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.
+
+#include "textflag.h"
+#include "funcdata.h"
+
+// bool armcas(int32 *val, int32 old, int32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	}else
+//		return 0;
+//
+// To implement ·cas in sys_$GOOS_arm.s
+// using the native instructions, use:
+//
+//	TEXT ·cas(SB),NOSPLIT,$0
+//		B	·armcas(SB)
+//
+TEXT ·armcas(SB),NOSPLIT,$0-13
+	MOVW	ptr+0(FP), R1
+	MOVW	old+4(FP), R2
+	MOVW	new+8(FP), R3
+casl:
+	LDREX	(R1), R0
+	CMP	R0, R2
+	BNE	casfail
+
+	MOVB	runtime·goarm(SB), R8
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISHST
+
+	STREX	R3, (R1), R0
+	CMP	$0, R0
+	BNE	casl
+	MOVW	$1, R0
+
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISH
+
+	MOVB	R0, ret+12(FP)
+	RET
+casfail:
+	MOVW	$0, R0
+	MOVB	R0, ret+12(FP)
+	RET
+
+// stubs
+
+TEXT ·Loadp(SB),NOSPLIT|NOFRAME,$0-8
+	B	·Load(SB)
+
+TEXT ·LoadAcq(SB),NOSPLIT|NOFRAME,$0-8
+	B	·Load(SB)
+
+TEXT ·LoadAcquintptr(SB),NOSPLIT|NOFRAME,$0-8
+	B 	·Load(SB)
+
+TEXT ·Casint32(SB),NOSPLIT,$0-13
+	B	·Cas(SB)
+
+TEXT ·Casint64(SB),NOSPLIT,$-4-21
+	B	·Cas64(SB)
+
+TEXT ·Casuintptr(SB),NOSPLIT,$0-13
+	B	·Cas(SB)
+
+TEXT ·Casp1(SB),NOSPLIT,$0-13
+	B	·Cas(SB)
+
+TEXT ·CasRel(SB),NOSPLIT,$0-13
+	B	·Cas(SB)
+
+TEXT ·Loadint32(SB),NOSPLIT,$0-8
+	B	·Load(SB)
+
+TEXT ·Loadint64(SB),NOSPLIT,$-4-12
+	B	·Load64(SB)
+
+TEXT ·Loaduintptr(SB),NOSPLIT,$0-8
+	B	·Load(SB)
+
+TEXT ·Loaduint(SB),NOSPLIT,$0-8
+	B	·Load(SB)
+
+TEXT ·Storeint32(SB),NOSPLIT,$0-8
+	B	·Store(SB)
+
+TEXT ·Storeint64(SB),NOSPLIT,$0-12
+	B	·Store64(SB)
+
+TEXT ·Storeuintptr(SB),NOSPLIT,$0-8
+	B	·Store(SB)
+
+TEXT ·StorepNoWB(SB),NOSPLIT,$0-8
+	B	·Store(SB)
+
+TEXT ·StoreRel(SB),NOSPLIT,$0-8
+	B	·Store(SB)
+
+TEXT ·StoreReluintptr(SB),NOSPLIT,$0-8
+	B	·Store(SB)
+
+TEXT ·Xaddint32(SB),NOSPLIT,$0-12
+	B	·Xadd(SB)
+
+TEXT ·Xaddint64(SB),NOSPLIT,$-4-20
+	B	·Xadd64(SB)
+
+TEXT ·Xadduintptr(SB),NOSPLIT,$0-12
+	B	·Xadd(SB)
+
+TEXT ·Xchgint32(SB),NOSPLIT,$0-12
+	B	·Xchg(SB)
+
+TEXT ·Xchgint64(SB),NOSPLIT,$-4-20
+	B	·Xchg64(SB)
+
+// 64-bit atomics
+// The native ARM implementations use LDREXD/STREXD, which are
+// available on ARMv6k or later. We use them only on ARMv7.
+// On older ARM, we use Go implementations which simulate 64-bit
+// atomics with locks.
+TEXT armCas64<>(SB),NOSPLIT,$0-21
+	// addr is already in R1
+	MOVW	old_lo+4(FP), R2
+	MOVW	old_hi+8(FP), R3
+	MOVW	new_lo+12(FP), R4
+	MOVW	new_hi+16(FP), R5
+cas64loop:
+	LDREXD	(R1), R6	// loads R6 and R7
+	CMP	R2, R6
+	BNE	cas64fail
+	CMP	R3, R7
+	BNE	cas64fail
+
+	DMB	MB_ISHST
+
+	STREXD	R4, (R1), R0	// stores R4 and R5
+	CMP	$0, R0
+	BNE	cas64loop
+	MOVW	$1, R0
+
+	DMB	MB_ISH
+
+	MOVBU	R0, swapped+20(FP)
+	RET
+cas64fail:
+	MOVW	$0, R0
+	MOVBU	R0, swapped+20(FP)
+	RET
+
+TEXT armXadd64<>(SB),NOSPLIT,$0-20
+	// addr is already in R1
+	MOVW	delta_lo+4(FP), R2
+	MOVW	delta_hi+8(FP), R3
+
+add64loop:
+	LDREXD	(R1), R4	// loads R4 and R5
+	ADD.S	R2, R4
+	ADC	R3, R5
+
+	DMB	MB_ISHST
+
+	STREXD	R4, (R1), R0	// stores R4 and R5
+	CMP	$0, R0
+	BNE	add64loop
+
+	DMB	MB_ISH
+
+	MOVW	R4, new_lo+12(FP)
+	MOVW	R5, new_hi+16(FP)
+	RET
+
+TEXT armXchg64<>(SB),NOSPLIT,$0-20
+	// addr is already in R1
+	MOVW	new_lo+4(FP), R2
+	MOVW	new_hi+8(FP), R3
+
+swap64loop:
+	LDREXD	(R1), R4	// loads R4 and R5
+
+	DMB	MB_ISHST
+
+	STREXD	R2, (R1), R0	// stores R2 and R3
+	CMP	$0, R0
+	BNE	swap64loop
+
+	DMB	MB_ISH
+
+	MOVW	R4, old_lo+12(FP)
+	MOVW	R5, old_hi+16(FP)
+	RET
+
+TEXT armLoad64<>(SB),NOSPLIT,$0-12
+	// addr is already in R1
+
+	LDREXD	(R1), R2	// loads R2 and R3
+	DMB	MB_ISH
+
+	MOVW	R2, val_lo+4(FP)
+	MOVW	R3, val_hi+8(FP)
+	RET
+
+TEXT armStore64<>(SB),NOSPLIT,$0-12
+	// addr is already in R1
+	MOVW	val_lo+4(FP), R2
+	MOVW	val_hi+8(FP), R3
+
+store64loop:
+	LDREXD	(R1), R4	// loads R4 and R5
+
+	DMB	MB_ISHST
+
+	STREXD	R2, (R1), R0	// stores R2 and R3
+	CMP	$0, R0
+	BNE	store64loop
+
+	DMB	MB_ISH
+	RET
+
+// The following functions all panic if their address argument isn't
+// 8-byte aligned. Since we're calling back into Go code to do this,
+// we have to cooperate with stack unwinding. In the normal case, the
+// functions tail-call into the appropriate implementation, which
+// means they must not open a frame. Hence, when they go down the
+// panic path, at that point they push the LR to create a real frame
+// (they don't need to pop it because panic won't return; however, we
+// do need to set the SP delta back).
+
+// Check if R1 is 8-byte aligned, panic if not.
+// Clobbers R2.
+#define CHECK_ALIGN \
+	AND.S	$7, R1, R2 \
+	BEQ 	4(PC) \
+	MOVW.W	R14, -4(R13) /* prepare a real frame */ \
+	BL	·panicUnaligned(SB) \
+	ADD	$4, R13 /* compensate SP delta */
+
+TEXT ·Cas64(SB),NOSPLIT,$-4-21
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+	CHECK_ALIGN
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	armCas64<>(SB)
+	JMP	·goCas64(SB)
+
+TEXT ·Xadd64(SB),NOSPLIT,$-4-20
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+	CHECK_ALIGN
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	armXadd64<>(SB)
+	JMP	·goXadd64(SB)
+
+TEXT ·Xchg64(SB),NOSPLIT,$-4-20
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+	CHECK_ALIGN
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	armXchg64<>(SB)
+	JMP	·goXchg64(SB)
+
+TEXT ·Load64(SB),NOSPLIT,$-4-12
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+	CHECK_ALIGN
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	armLoad64<>(SB)
+	JMP	·goLoad64(SB)
+
+TEXT ·Store64(SB),NOSPLIT,$-4-12
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+	CHECK_ALIGN
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	armStore64<>(SB)
+	JMP	·goStore64(SB)
diff --git a/src/runtime/internal/atomic/atomic_arm64.go b/src/runtime/internal/atomic/atomic_arm64.go
new file mode 100644
index 0000000..459fb99
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_arm64.go
@@ -0,0 +1,94 @@
+// Copyright 2015 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 arm64
+
+package atomic
+
+import (
+	"internal/cpu"
+	"unsafe"
+)
+
+const (
+	offsetARM64HasATOMICS = unsafe.Offsetof(cpu.ARM64.HasATOMICS)
+)
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(addr *uint32) uint32
+
+//go:noescape
+func LoadAcq64(ptr *uint64) uint64
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreRel64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
diff --git a/src/runtime/internal/atomic/atomic_arm64.s b/src/runtime/internal/atomic/atomic_arm64.s
new file mode 100644
index 0000000..5f77d92
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_arm64.s
@@ -0,0 +1,333 @@
+// Copyright 2014 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.
+
+#include "go_asm.h"
+#include "textflag.h"
+
+TEXT ·Casint32(SB), NOSPLIT, $0-17
+	B	·Cas(SB)
+
+TEXT ·Casint64(SB), NOSPLIT, $0-25
+	B	·Cas64(SB)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	B	·Cas64(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	B	·Cas(SB)
+
+TEXT ·Loadint32(SB), NOSPLIT, $0-12
+	B	·Load(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	B	·Load64(SB)
+
+TEXT ·Loaduintptr(SB), NOSPLIT, $0-16
+	B	·Load64(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT, $0-16
+	B	·Load64(SB)
+
+TEXT ·Storeint32(SB), NOSPLIT, $0-12
+	B	·Store(SB)
+
+TEXT ·Storeint64(SB), NOSPLIT, $0-16
+	B	·Store64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	B	·Store64(SB)
+
+TEXT ·Xaddint32(SB), NOSPLIT, $0-20
+	B	·Xadd(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	B	·Xadd64(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	B	·Xadd64(SB)
+
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	B ·Cas64(SB)
+
+// uint32 ·Load(uint32 volatile* addr)
+TEXT ·Load(SB),NOSPLIT,$0-12
+	MOVD	ptr+0(FP), R0
+	LDARW	(R0), R0
+	MOVW	R0, ret+8(FP)
+	RET
+
+// uint8 ·Load8(uint8 volatile* addr)
+TEXT ·Load8(SB),NOSPLIT,$0-9
+	MOVD	ptr+0(FP), R0
+	LDARB	(R0), R0
+	MOVB	R0, ret+8(FP)
+	RET
+
+// uint64 ·Load64(uint64 volatile* addr)
+TEXT ·Load64(SB),NOSPLIT,$0-16
+	MOVD	ptr+0(FP), R0
+	LDAR	(R0), R0
+	MOVD	R0, ret+8(FP)
+	RET
+
+// void *·Loadp(void *volatile *addr)
+TEXT ·Loadp(SB),NOSPLIT,$0-16
+	MOVD	ptr+0(FP), R0
+	LDAR	(R0), R0
+	MOVD	R0, ret+8(FP)
+	RET
+
+// uint32 ·LoadAcq(uint32 volatile* addr)
+TEXT ·LoadAcq(SB),NOSPLIT,$0-12
+	B	·Load(SB)
+
+// uint64 ·LoadAcquintptr(uint64 volatile* addr)
+TEXT ·LoadAcq64(SB),NOSPLIT,$0-16
+	B	·Load64(SB)
+
+// uintptr ·LoadAcq64(uintptr volatile* addr)
+TEXT ·LoadAcquintptr(SB),NOSPLIT,$0-16
+	B	·Load64(SB)
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	B	·Store64(SB)
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	B	·Store(SB)
+
+TEXT ·StoreRel64(SB), NOSPLIT, $0-16
+	B	·Store64(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	B	·Store64(SB)
+
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R0
+	MOVW	val+8(FP), R1
+	STLRW	R1, (R0)
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R0
+	MOVB	val+8(FP), R1
+	STLRB	R1, (R0)
+	RET
+
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOVD	ptr+0(FP), R0
+	MOVD	val+8(FP), R1
+	STLR	R1, (R0)
+	RET
+
+// uint32 Xchg(ptr *uint32, new uint32)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R0
+	MOVW	new+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	SWPALW	R1, (R0), R2
+	MOVW	R2, ret+16(FP)
+	RET
+load_store_loop:
+	LDAXRW	(R0), R2
+	STLXRW	R1, (R0), R3
+	CBNZ	R3, load_store_loop
+	MOVW	R2, ret+16(FP)
+	RET
+
+// uint64 Xchg64(ptr *uint64, new uint64)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R0
+	MOVD	new+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	SWPALD	R1, (R0), R2
+	MOVD	R2, ret+16(FP)
+	RET
+load_store_loop:
+	LDAXR	(R0), R2
+	STLXR	R1, (R0), R3
+	CBNZ	R3, load_store_loop
+	MOVD	R2, ret+16(FP)
+	RET
+
+// bool Cas(uint32 *ptr, uint32 old, uint32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOVD	ptr+0(FP), R0
+	MOVW	old+8(FP), R1
+	MOVW	new+12(FP), R2
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	MOVD	R1, R3
+	CASALW	R3, (R0), R2
+	CMP 	R1, R3
+	CSET	EQ, R0
+	MOVB	R0, ret+16(FP)
+	RET
+load_store_loop:
+	LDAXRW	(R0), R3
+	CMPW	R1, R3
+	BNE	ok
+	STLXRW	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+ok:
+	CSET	EQ, R0
+	MOVB	R0, ret+16(FP)
+	RET
+
+// bool ·Cas64(uint64 *ptr, uint64 old, uint64 new)
+// Atomically:
+//      if(*val == old){
+//              *val = new;
+//              return 1;
+//      } else {
+//              return 0;
+//      }
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVD	ptr+0(FP), R0
+	MOVD	old+8(FP), R1
+	MOVD	new+16(FP), R2
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	MOVD	R1, R3
+	CASALD	R3, (R0), R2
+	CMP 	R1, R3
+	CSET	EQ, R0
+	MOVB	R0, ret+24(FP)
+	RET
+load_store_loop:
+	LDAXR	(R0), R3
+	CMP	R1, R3
+	BNE	ok
+	STLXR	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+ok:
+	CSET	EQ, R0
+	MOVB	R0, ret+24(FP)
+	RET
+
+// uint32 xadd(uint32 volatile *ptr, int32 delta)
+// Atomically:
+//      *val += delta;
+//      return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R0
+	MOVW	delta+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	LDADDALW	R1, (R0), R2
+	ADD 	R1, R2
+	MOVW	R2, ret+16(FP)
+	RET
+load_store_loop:
+	LDAXRW	(R0), R2
+	ADDW	R2, R1, R2
+	STLXRW	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	MOVW	R2, ret+16(FP)
+	RET
+
+// uint64 Xadd64(uint64 volatile *ptr, int64 delta)
+// Atomically:
+//      *val += delta;
+//      return *val;
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R0
+	MOVD	delta+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	LDADDALD	R1, (R0), R2
+	ADD 	R1, R2
+	MOVD	R2, ret+16(FP)
+	RET
+load_store_loop:
+	LDAXR	(R0), R2
+	ADD	R2, R1, R2
+	STLXR	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	MOVD	R2, ret+16(FP)
+	RET
+
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	B	·Xchg(SB)
+
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	B	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	B	·Xchg64(SB)
+
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R0
+	MOVB	val+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	MVN 	R1, R2
+	LDCLRALB	R2, (R0), R3
+	RET
+load_store_loop:
+	LDAXRB	(R0), R2
+	AND	R1, R2
+	STLXRB	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	RET
+
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R0
+	MOVB	val+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	LDORALB	R1, (R0), R2
+	RET
+load_store_loop:
+	LDAXRB	(R0), R2
+	ORR	R1, R2
+	STLXRB	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R0
+	MOVW	val+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	MVN 	R1, R2
+	LDCLRALW	R2, (R0), R3
+	RET
+load_store_loop:
+	LDAXRW	(R0), R2
+	AND	R1, R2
+	STLXRW	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R0
+	MOVW	val+8(FP), R1
+	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
+	CBZ 	R4, load_store_loop
+	LDORALW	R1, (R0), R2
+	RET
+load_store_loop:
+	LDAXRW	(R0), R2
+	ORR	R1, R2
+	STLXRW	R2, (R0), R3
+	CBNZ	R3, load_store_loop
+	RET
diff --git a/src/runtime/internal/atomic/atomic_loong64.go b/src/runtime/internal/atomic/atomic_loong64.go
new file mode 100644
index 0000000..908a7d6
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_loong64.go
@@ -0,0 +1,83 @@
+// Copyright 2022 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 loong64
+
+package atomic
+
+import "unsafe"
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(ptr *uint32) uint32
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
diff --git a/src/runtime/internal/atomic/atomic_loong64.s b/src/runtime/internal/atomic/atomic_loong64.s
new file mode 100644
index 0000000..bfb6c7e
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_loong64.s
@@ -0,0 +1,299 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// bool cas(uint32 *ptr, uint32 old, uint32 new)
+// Atomically:
+//	if(*ptr == old){
+//		*ptr = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOVV	ptr+0(FP), R4
+	MOVW	old+8(FP), R5
+	MOVW	new+12(FP), R6
+	DBAR
+cas_again:
+	MOVV	R6, R7
+	LL	(R4), R8
+	BNE	R5, R8, cas_fail
+	SC	R7, (R4)
+	BEQ	R7, cas_again
+	MOVV	$1, R4
+	MOVB	R4, ret+16(FP)
+	DBAR
+	RET
+cas_fail:
+	MOVV	$0, R4
+	JMP	-4(PC)
+
+// bool	cas64(uint64 *ptr, uint64 old, uint64 new)
+// Atomically:
+//	if(*ptr == old){
+//		*ptr = new;
+//		return 1;
+//	} else {
+//		return 0;
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVV	ptr+0(FP), R4
+	MOVV	old+8(FP), R5
+	MOVV	new+16(FP), R6
+	DBAR
+cas64_again:
+	MOVV	R6, R7
+	LLV	(R4), R8
+	BNE	R5, R8, cas64_fail
+	SCV	R7, (R4)
+	BEQ	R7, cas64_again
+	MOVV	$1, R4
+	MOVB	R4, ret+24(FP)
+	DBAR
+	RET
+cas64_fail:
+	MOVV	$0, R4
+	JMP	-4(PC)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+TEXT ·Loaduintptr(SB),  NOSPLIT|NOFRAME, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT|NOFRAME, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+// bool casp(void **val, void *old, void *new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	JMP runtime∕internal∕atomic·Cas64(SB)
+
+// uint32 xadd(uint32 volatile *ptr, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVV	ptr+0(FP), R4
+	MOVW	delta+8(FP), R5
+	DBAR
+	LL	(R4), R6
+	ADDU	R6, R5, R7
+	MOVV	R7, R6
+	SC	R7, (R4)
+	BEQ	R7, -4(PC)
+	MOVW	R6, ret+16(FP)
+	DBAR
+	RET
+
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVV	ptr+0(FP), R4
+	MOVV	delta+8(FP), R5
+	DBAR
+	LLV	(R4), R6
+	ADDVU	R6, R5, R7
+	MOVV	R7, R6
+	SCV	R7, (R4)
+	BEQ	R7, -4(PC)
+	MOVV	R6, ret+16(FP)
+	DBAR
+	RET
+
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVV	ptr+0(FP), R4
+	MOVW	new+8(FP), R5
+
+	DBAR
+	MOVV	R5, R6
+	LL	(R4), R7
+	SC	R6, (R4)
+	BEQ	R6, -3(PC)
+	MOVW	R7, ret+16(FP)
+	DBAR
+	RET
+
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVV	ptr+0(FP), R4
+	MOVV	new+8(FP), R5
+
+	DBAR
+	MOVV	R5, R6
+	LLV	(R4), R7
+	SCV	R6, (R4)
+	BEQ	R6, -3(PC)
+	MOVV	R7, ret+16(FP)
+	DBAR
+	RET
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	JMP     ·Store64(SB)
+
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R4
+	MOVW	val+8(FP), R5
+	DBAR
+	MOVW	R5, 0(R4)
+	DBAR
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R4
+	MOVB	val+8(FP), R5
+	DBAR
+	MOVB	R5, 0(R4)
+	DBAR
+	RET
+
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOVV	ptr+0(FP), R4
+	MOVV	val+8(FP), R5
+	DBAR
+	MOVV	R5, 0(R4)
+	DBAR
+	RET
+
+// void	Or8(byte volatile*, byte);
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R4
+	MOVBU	val+8(FP), R5
+	// Align ptr down to 4 bytes so we can use 32-bit load/store.
+	MOVV	$~3, R6
+	AND	R4, R6
+	// R7 = ((ptr & 3) * 8)
+	AND	$3, R4, R7
+	SLLV	$3, R7
+	// Shift val for aligned ptr. R5 = val << R4
+	SLLV	R7, R5
+
+	DBAR
+	LL	(R6), R7
+	OR	R5, R7
+	SC	R7, (R6)
+	BEQ	R7, -4(PC)
+	DBAR
+	RET
+
+// void	And8(byte volatile*, byte);
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R4
+	MOVBU	val+8(FP), R5
+	// Align ptr down to 4 bytes so we can use 32-bit load/store.
+	MOVV	$~3, R6
+	AND	R4, R6
+	// R7 = ((ptr & 3) * 8)
+	AND	$3, R4, R7
+	SLLV	$3, R7
+	// Shift val for aligned ptr. R5 = val << R7 | ^(0xFF << R7)
+	MOVV	$0xFF, R8
+	SLLV	R7, R5
+	SLLV	R7, R8
+	NOR	R0, R8
+	OR	R8, R5
+
+	DBAR
+	LL	(R6), R7
+	AND	R5, R7
+	SC	R7, (R6)
+	BEQ	R7, -4(PC)
+	DBAR
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R4
+	MOVW	val+8(FP), R5
+	DBAR
+	LL	(R4), R6
+	OR	R5, R6
+	SC	R6, (R4)
+	BEQ	R6, -4(PC)
+	DBAR
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R4
+	MOVW	val+8(FP), R5
+	DBAR
+	LL	(R4), R6
+	AND	R5, R6
+	SC	R6, (R4)
+	BEQ	R6, -4(PC)
+	DBAR
+	RET
+
+// uint32 runtime∕internal∕atomic·Load(uint32 volatile* ptr)
+TEXT ·Load(SB),NOSPLIT|NOFRAME,$0-12
+	MOVV	ptr+0(FP), R19
+	DBAR
+	MOVWU	0(R19), R19
+	DBAR
+	MOVW	R19, ret+8(FP)
+	RET
+
+// uint8 runtime∕internal∕atomic·Load8(uint8 volatile* ptr)
+TEXT ·Load8(SB),NOSPLIT|NOFRAME,$0-9
+	MOVV	ptr+0(FP), R19
+	DBAR
+	MOVBU	0(R19), R19
+	DBAR
+	MOVB	R19, ret+8(FP)
+	RET
+
+// uint64 runtime∕internal∕atomic·Load64(uint64 volatile* ptr)
+TEXT ·Load64(SB),NOSPLIT|NOFRAME,$0-16
+	MOVV	ptr+0(FP), R19
+	DBAR
+	MOVV	0(R19), R19
+	DBAR
+	MOVV	R19, ret+8(FP)
+	RET
+
+// void *runtime∕internal∕atomic·Loadp(void *volatile *ptr)
+TEXT ·Loadp(SB),NOSPLIT|NOFRAME,$0-16
+	MOVV	ptr+0(FP), R19
+	DBAR
+	MOVV	0(R19), R19
+	DBAR
+	MOVV	R19, ret+8(FP)
+	RET
+
+// uint32 runtime∕internal∕atomic·LoadAcq(uint32 volatile* ptr)
+TEXT ·LoadAcq(SB),NOSPLIT|NOFRAME,$0-12
+	JMP	atomic·Load(SB)
+
+// uintptr ·LoadAcquintptr(uintptr volatile* ptr)
+TEXT ·LoadAcquintptr(SB),NOSPLIT|NOFRAME,$0-16
+	JMP     atomic·Load64(SB)
+
diff --git a/src/runtime/internal/atomic/atomic_mips64x.go b/src/runtime/internal/atomic/atomic_mips64x.go
new file mode 100644
index 0000000..1e12b83
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_mips64x.go
@@ -0,0 +1,89 @@
+// Copyright 2015 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 mips64 || mips64le
+
+package atomic
+
+import "unsafe"
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(ptr *uint32) uint32
+
+//go:noescape
+func LoadAcq64(ptr *uint64) uint64
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreRel64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
diff --git a/src/runtime/internal/atomic/atomic_mips64x.s b/src/runtime/internal/atomic/atomic_mips64x.s
new file mode 100644
index 0000000..b4411d8
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_mips64x.s
@@ -0,0 +1,359 @@
+// Copyright 2015 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 mips64 || mips64le
+
+#include "textflag.h"
+
+#define SYNC	WORD $0xf
+
+// bool cas(uint32 *ptr, uint32 old, uint32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOVV	ptr+0(FP), R1
+	MOVW	old+8(FP), R2
+	MOVW	new+12(FP), R5
+	SYNC
+cas_again:
+	MOVV	R5, R3
+	LL	(R1), R4
+	BNE	R2, R4, cas_fail
+	SC	R3, (R1)
+	BEQ	R3, cas_again
+	MOVV	$1, R1
+	MOVB	R1, ret+16(FP)
+	SYNC
+	RET
+cas_fail:
+	MOVV	$0, R1
+	JMP	-4(PC)
+
+// bool	cas64(uint64 *ptr, uint64 old, uint64 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		return 0;
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVV	ptr+0(FP), R1
+	MOVV	old+8(FP), R2
+	MOVV	new+16(FP), R5
+	SYNC
+cas64_again:
+	MOVV	R5, R3
+	LLV	(R1), R4
+	BNE	R2, R4, cas64_fail
+	SCV	R3, (R1)
+	BEQ	R3, cas64_again
+	MOVV	$1, R1
+	MOVB	R1, ret+24(FP)
+	SYNC
+	RET
+cas64_fail:
+	MOVV	$0, R1
+	JMP	-4(PC)
+
+TEXT ·Casint32(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+TEXT ·Casint64(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+TEXT ·Loaduintptr(SB),  NOSPLIT|NOFRAME, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT|NOFRAME, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Storeint32(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·Storeint64(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+TEXT ·Loadint32(SB), NOSPLIT, $0-12
+	JMP	·Load(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	JMP	·Load64(SB)
+
+TEXT ·Xaddint32(SB), NOSPLIT, $0-20
+	JMP	·Xadd(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+// bool casp(void **val, void *old, void *new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	JMP ·Cas64(SB)
+
+// uint32 xadd(uint32 volatile *ptr, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVV	ptr+0(FP), R2
+	MOVW	delta+8(FP), R3
+	SYNC
+	LL	(R2), R1
+	ADDU	R1, R3, R4
+	MOVV	R4, R1
+	SC	R4, (R2)
+	BEQ	R4, -4(PC)
+	MOVW	R1, ret+16(FP)
+	SYNC
+	RET
+
+// uint64 Xadd64(uint64 volatile *ptr, int64 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVV	ptr+0(FP), R2
+	MOVV	delta+8(FP), R3
+	SYNC
+	LLV	(R2), R1
+	ADDVU	R1, R3, R4
+	MOVV	R4, R1
+	SCV	R4, (R2)
+	BEQ	R4, -4(PC)
+	MOVV	R1, ret+16(FP)
+	SYNC
+	RET
+
+// uint32 Xchg(ptr *uint32, new uint32)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVV	ptr+0(FP), R2
+	MOVW	new+8(FP), R5
+
+	SYNC
+	MOVV	R5, R3
+	LL	(R2), R1
+	SC	R3, (R2)
+	BEQ	R3, -3(PC)
+	MOVW	R1, ret+16(FP)
+	SYNC
+	RET
+
+// uint64 Xchg64(ptr *uint64, new uint64)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVV	ptr+0(FP), R2
+	MOVV	new+8(FP), R5
+
+	SYNC
+	MOVV	R5, R3
+	LLV	(R2), R1
+	SCV	R3, (R2)
+	BEQ	R3, -3(PC)
+	MOVV	R1, ret+16(FP)
+	SYNC
+	RET
+
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	JMP	·Xchg(SB)
+
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·StoreRel64(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R1
+	MOVW	val+8(FP), R2
+	SYNC
+	MOVW	R2, 0(R1)
+	SYNC
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R1
+	MOVB	val+8(FP), R2
+	SYNC
+	MOVB	R2, 0(R1)
+	SYNC
+	RET
+
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOVV	ptr+0(FP), R1
+	MOVV	val+8(FP), R2
+	SYNC
+	MOVV	R2, 0(R1)
+	SYNC
+	RET
+
+// void	Or8(byte volatile*, byte);
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R1
+	MOVBU	val+8(FP), R2
+	// Align ptr down to 4 bytes so we can use 32-bit load/store.
+	MOVV	$~3, R3
+	AND	R1, R3
+	// Compute val shift.
+#ifdef GOARCH_mips64
+	// Big endian.  ptr = ptr ^ 3
+	XOR	$3, R1
+#endif
+	// R4 = ((ptr & 3) * 8)
+	AND	$3, R1, R4
+	SLLV	$3, R4
+	// Shift val for aligned ptr. R2 = val << R4
+	SLLV	R4, R2
+
+	SYNC
+	LL	(R3), R4
+	OR	R2, R4
+	SC	R4, (R3)
+	BEQ	R4, -4(PC)
+	SYNC
+	RET
+
+// void	And8(byte volatile*, byte);
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVV	ptr+0(FP), R1
+	MOVBU	val+8(FP), R2
+	// Align ptr down to 4 bytes so we can use 32-bit load/store.
+	MOVV	$~3, R3
+	AND	R1, R3
+	// Compute val shift.
+#ifdef GOARCH_mips64
+	// Big endian.  ptr = ptr ^ 3
+	XOR	$3, R1
+#endif
+	// R4 = ((ptr & 3) * 8)
+	AND	$3, R1, R4
+	SLLV	$3, R4
+	// Shift val for aligned ptr. R2 = val << R4 | ^(0xFF << R4)
+	MOVV	$0xFF, R5
+	SLLV	R4, R2
+	SLLV	R4, R5
+	NOR	R0, R5
+	OR	R5, R2
+
+	SYNC
+	LL	(R3), R4
+	AND	R2, R4
+	SC	R4, (R3)
+	BEQ	R4, -4(PC)
+	SYNC
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R1
+	MOVW	val+8(FP), R2
+
+	SYNC
+	LL	(R1), R3
+	OR	R2, R3
+	SC	R3, (R1)
+	BEQ	R3, -4(PC)
+	SYNC
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVV	ptr+0(FP), R1
+	MOVW	val+8(FP), R2
+
+	SYNC
+	LL	(R1), R3
+	AND	R2, R3
+	SC	R3, (R1)
+	BEQ	R3, -4(PC)
+	SYNC
+	RET
+
+// uint32 ·Load(uint32 volatile* ptr)
+TEXT ·Load(SB),NOSPLIT|NOFRAME,$0-12
+	MOVV	ptr+0(FP), R1
+	SYNC
+	MOVWU	0(R1), R1
+	SYNC
+	MOVW	R1, ret+8(FP)
+	RET
+
+// uint8 ·Load8(uint8 volatile* ptr)
+TEXT ·Load8(SB),NOSPLIT|NOFRAME,$0-9
+	MOVV	ptr+0(FP), R1
+	SYNC
+	MOVBU	0(R1), R1
+	SYNC
+	MOVB	R1, ret+8(FP)
+	RET
+
+// uint64 ·Load64(uint64 volatile* ptr)
+TEXT ·Load64(SB),NOSPLIT|NOFRAME,$0-16
+	MOVV	ptr+0(FP), R1
+	SYNC
+	MOVV	0(R1), R1
+	SYNC
+	MOVV	R1, ret+8(FP)
+	RET
+
+// void *·Loadp(void *volatile *ptr)
+TEXT ·Loadp(SB),NOSPLIT|NOFRAME,$0-16
+	MOVV	ptr+0(FP), R1
+	SYNC
+	MOVV	0(R1), R1
+	SYNC
+	MOVV	R1, ret+8(FP)
+	RET
+
+// uint32 ·LoadAcq(uint32 volatile* ptr)
+TEXT ·LoadAcq(SB),NOSPLIT|NOFRAME,$0-12
+	JMP	atomic·Load(SB)
+
+// uint64 ·LoadAcq64(uint64 volatile* ptr)
+TEXT ·LoadAcq64(SB),NOSPLIT|NOFRAME,$0-16
+	JMP	atomic·Load64(SB)
+
+// uintptr ·LoadAcquintptr(uintptr volatile* ptr)
+TEXT ·LoadAcquintptr(SB),NOSPLIT|NOFRAME,$0-16
+	JMP	atomic·Load64(SB)
diff --git a/src/runtime/internal/atomic/atomic_mipsx.go b/src/runtime/internal/atomic/atomic_mipsx.go
new file mode 100644
index 0000000..5dd15a0
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_mipsx.go
@@ -0,0 +1,167 @@
+// Copyright 2016 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 mips || mipsle
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Xadd64
+//go:linkname Xchg64
+//go:linkname Cas64
+//go:linkname Load64
+//go:linkname Store64
+
+package atomic
+
+import (
+	"internal/cpu"
+	"unsafe"
+)
+
+// TODO implement lock striping
+var lock struct {
+	state uint32
+	pad   [cpu.CacheLinePadSize - 4]byte
+}
+
+//go:noescape
+func spinLock(state *uint32)
+
+//go:noescape
+func spinUnlock(state *uint32)
+
+//go:nosplit
+func lockAndCheck(addr *uint64) {
+	// ensure 8-byte alignment
+	if uintptr(unsafe.Pointer(addr))&7 != 0 {
+		panicUnaligned()
+	}
+	// force dereference before taking lock
+	_ = *addr
+
+	spinLock(&lock.state)
+}
+
+//go:nosplit
+func unlock() {
+	spinUnlock(&lock.state)
+}
+
+//go:nosplit
+func unlockNoFence() {
+	lock.state = 0
+}
+
+//go:nosplit
+func Xadd64(addr *uint64, delta int64) (new uint64) {
+	lockAndCheck(addr)
+
+	new = *addr + uint64(delta)
+	*addr = new
+
+	unlock()
+	return
+}
+
+//go:nosplit
+func Xchg64(addr *uint64, new uint64) (old uint64) {
+	lockAndCheck(addr)
+
+	old = *addr
+	*addr = new
+
+	unlock()
+	return
+}
+
+//go:nosplit
+func Cas64(addr *uint64, old, new uint64) (swapped bool) {
+	lockAndCheck(addr)
+
+	if (*addr) == old {
+		*addr = new
+		unlock()
+		return true
+	}
+
+	unlockNoFence()
+	return false
+}
+
+//go:nosplit
+func Load64(addr *uint64) (val uint64) {
+	lockAndCheck(addr)
+
+	val = *addr
+
+	unlock()
+	return
+}
+
+//go:nosplit
+func Store64(addr *uint64, val uint64) {
+	lockAndCheck(addr)
+
+	*addr = val
+
+	unlock()
+	return
+}
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(ptr *uint32) uint32
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
+
+//go:noescape
+func CasRel(addr *uint32, old, new uint32) bool
diff --git a/src/runtime/internal/atomic/atomic_mipsx.s b/src/runtime/internal/atomic/atomic_mipsx.s
new file mode 100644
index 0000000..390e9ce
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_mipsx.s
@@ -0,0 +1,261 @@
+// Copyright 2016 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 mips || mipsle
+
+#include "textflag.h"
+
+// bool Cas(int32 *val, int32 old, int32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB),NOSPLIT,$0-13
+	MOVW	ptr+0(FP), R1
+	MOVW	old+4(FP), R2
+	MOVW	new+8(FP), R5
+	SYNC
+try_cas:
+	MOVW	R5, R3
+	LL	(R1), R4	// R4 = *R1
+	BNE	R2, R4, cas_fail
+	SC	R3, (R1)	// *R1 = R3
+	BEQ	R3, try_cas
+	SYNC
+	MOVB	R3, ret+12(FP)
+	RET
+cas_fail:
+	MOVB	R0, ret+12(FP)
+	RET
+
+TEXT ·Store(SB),NOSPLIT,$0-8
+	MOVW	ptr+0(FP), R1
+	MOVW	val+4(FP), R2
+	SYNC
+	MOVW	R2, 0(R1)
+	SYNC
+	RET
+
+TEXT ·Store8(SB),NOSPLIT,$0-5
+	MOVW	ptr+0(FP), R1
+	MOVB	val+4(FP), R2
+	SYNC
+	MOVB	R2, 0(R1)
+	SYNC
+	RET
+
+TEXT ·Load(SB),NOSPLIT,$0-8
+	MOVW	ptr+0(FP), R1
+	SYNC
+	MOVW	0(R1), R1
+	SYNC
+	MOVW	R1, ret+4(FP)
+	RET
+
+TEXT ·Load8(SB),NOSPLIT,$0-5
+	MOVW	ptr+0(FP), R1
+	SYNC
+	MOVB	0(R1), R1
+	SYNC
+	MOVB	R1, ret+4(FP)
+	RET
+
+// uint32 Xadd(uint32 volatile *val, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB),NOSPLIT,$0-12
+	MOVW	ptr+0(FP), R2
+	MOVW	delta+4(FP), R3
+	SYNC
+try_xadd:
+	LL	(R2), R1	// R1 = *R2
+	ADDU	R1, R3, R4
+	MOVW	R4, R1
+	SC	R4, (R2)	// *R2 = R4
+	BEQ	R4, try_xadd
+	SYNC
+	MOVW	R1, ret+8(FP)
+	RET
+
+// uint32 Xchg(ptr *uint32, new uint32)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg(SB),NOSPLIT,$0-12
+	MOVW	ptr+0(FP), R2
+	MOVW	new+4(FP), R5
+	SYNC
+try_xchg:
+	MOVW	R5, R3
+	LL	(R2), R1	// R1 = *R2
+	SC	R3, (R2)	// *R2 = R3
+	BEQ	R3, try_xchg
+	SYNC
+	MOVW	R1, ret+8(FP)
+	RET
+
+TEXT ·Casint32(SB),NOSPLIT,$0-13
+	JMP	·Cas(SB)
+
+TEXT ·Casint64(SB),NOSPLIT,$0-21
+	JMP	·Cas64(SB)
+
+TEXT ·Casuintptr(SB),NOSPLIT,$0-13
+	JMP	·Cas(SB)
+
+TEXT ·CasRel(SB),NOSPLIT,$0-13
+	JMP	·Cas(SB)
+
+TEXT ·Loaduintptr(SB),NOSPLIT,$0-8
+	JMP	·Load(SB)
+
+TEXT ·Loaduint(SB),NOSPLIT,$0-8
+	JMP	·Load(SB)
+
+TEXT ·Loadp(SB),NOSPLIT,$-0-8
+	JMP	·Load(SB)
+
+TEXT ·Storeint32(SB),NOSPLIT,$0-8
+	JMP	·Store(SB)
+
+TEXT ·Storeint64(SB),NOSPLIT,$0-12
+	JMP	·Store64(SB)
+
+TEXT ·Storeuintptr(SB),NOSPLIT,$0-8
+	JMP	·Store(SB)
+
+TEXT ·Xadduintptr(SB),NOSPLIT,$0-12
+	JMP	·Xadd(SB)
+
+TEXT ·Loadint32(SB),NOSPLIT,$0-8
+	JMP	·Load(SB)
+
+TEXT ·Loadint64(SB),NOSPLIT,$0-12
+	JMP	·Load64(SB)
+
+TEXT ·Xaddint32(SB),NOSPLIT,$0-12
+	JMP	·Xadd(SB)
+
+TEXT ·Xaddint64(SB),NOSPLIT,$0-20
+	JMP	·Xadd64(SB)
+
+TEXT ·Casp1(SB),NOSPLIT,$0-13
+	JMP	·Cas(SB)
+
+TEXT ·Xchgint32(SB),NOSPLIT,$0-12
+	JMP	·Xchg(SB)
+
+TEXT ·Xchgint64(SB),NOSPLIT,$0-20
+	JMP	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB),NOSPLIT,$0-12
+	JMP	·Xchg(SB)
+
+TEXT ·StorepNoWB(SB),NOSPLIT,$0-8
+	JMP	·Store(SB)
+
+TEXT ·StoreRel(SB),NOSPLIT,$0-8
+	JMP	·Store(SB)
+
+TEXT ·StoreReluintptr(SB),NOSPLIT,$0-8
+	JMP	·Store(SB)
+
+// void	Or8(byte volatile*, byte);
+TEXT ·Or8(SB),NOSPLIT,$0-5
+	MOVW	ptr+0(FP), R1
+	MOVBU	val+4(FP), R2
+	MOVW	$~3, R3	// Align ptr down to 4 bytes so we can use 32-bit load/store.
+	AND	R1, R3
+#ifdef GOARCH_mips
+	// Big endian.  ptr = ptr ^ 3
+	XOR	$3, R1
+#endif
+	AND	$3, R1, R4	// R4 = ((ptr & 3) * 8)
+	SLL	$3, R4
+	SLL	R4, R2, R2	// Shift val for aligned ptr. R2 = val << R4
+	SYNC
+try_or8:
+	LL	(R3), R4	// R4 = *R3
+	OR	R2, R4
+	SC	R4, (R3)	// *R3 = R4
+	BEQ	R4, try_or8
+	SYNC
+	RET
+
+// void	And8(byte volatile*, byte);
+TEXT ·And8(SB),NOSPLIT,$0-5
+	MOVW	ptr+0(FP), R1
+	MOVBU	val+4(FP), R2
+	MOVW	$~3, R3
+	AND	R1, R3
+#ifdef GOARCH_mips
+	// Big endian.  ptr = ptr ^ 3
+	XOR	$3, R1
+#endif
+	AND	$3, R1, R4	// R4 = ((ptr & 3) * 8)
+	SLL	$3, R4
+	MOVW	$0xFF, R5
+	SLL	R4, R2
+	SLL	R4, R5
+	NOR	R0, R5
+	OR	R5, R2	// Shift val for aligned ptr. R2 = val << R4 | ^(0xFF << R4)
+	SYNC
+try_and8:
+	LL	(R3), R4	// R4 = *R3
+	AND	R2, R4
+	SC	R4, (R3)	// *R3 = R4
+	BEQ	R4, try_and8
+	SYNC
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-8
+	MOVW	ptr+0(FP), R1
+	MOVW	val+4(FP), R2
+
+	SYNC
+	LL	(R1), R3
+	OR	R2, R3
+	SC	R3, (R1)
+	BEQ	R3, -4(PC)
+	SYNC
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-8
+	MOVW	ptr+0(FP), R1
+	MOVW	val+4(FP), R2
+
+	SYNC
+	LL	(R1), R3
+	AND	R2, R3
+	SC	R3, (R1)
+	BEQ	R3, -4(PC)
+	SYNC
+	RET
+
+TEXT ·spinLock(SB),NOSPLIT,$0-4
+	MOVW	state+0(FP), R1
+	MOVW	$1, R2
+	SYNC
+try_lock:
+	MOVW	R2, R3
+check_again:
+	LL	(R1), R4
+	BNE	R4, check_again
+	SC	R3, (R1)
+	BEQ	R3, try_lock
+	SYNC
+	RET
+
+TEXT ·spinUnlock(SB),NOSPLIT,$0-4
+	MOVW	state+0(FP), R1
+	SYNC
+	MOVW	R0, (R1)
+	SYNC
+	RET
diff --git a/src/runtime/internal/atomic/atomic_ppc64x.go b/src/runtime/internal/atomic/atomic_ppc64x.go
new file mode 100644
index 0000000..998d16e
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_ppc64x.go
@@ -0,0 +1,89 @@
+// Copyright 2014 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 ppc64 || ppc64le
+
+package atomic
+
+import "unsafe"
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(ptr *uint32) uint32
+
+//go:noescape
+func LoadAcq64(ptr *uint64) uint64
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreRel64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
diff --git a/src/runtime/internal/atomic/atomic_ppc64x.s b/src/runtime/internal/atomic/atomic_ppc64x.s
new file mode 100644
index 0000000..04f0ead
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_ppc64x.s
@@ -0,0 +1,362 @@
+// Copyright 2014 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 ppc64 || ppc64le
+
+#include "textflag.h"
+
+// For more details about how various memory models are
+// enforced on POWER, the following paper provides more
+// details about how they enforce C/C++ like models. This
+// gives context about why the strange looking code
+// sequences below work.
+//
+// http://www.rdrop.com/users/paulmck/scalability/paper/N2745r.2011.03.04a.html
+
+// uint32 ·Load(uint32 volatile* ptr)
+TEXT ·Load(SB),NOSPLIT|NOFRAME,$-8-12
+	MOVD	ptr+0(FP), R3
+	SYNC
+	MOVWZ	0(R3), R3
+	CMPW	R3, R3, CR7
+	BC	4, 30, 1(PC) // bne- cr7,0x4
+	ISYNC
+	MOVW	R3, ret+8(FP)
+	RET
+
+// uint8 ·Load8(uint8 volatile* ptr)
+TEXT ·Load8(SB),NOSPLIT|NOFRAME,$-8-9
+	MOVD	ptr+0(FP), R3
+	SYNC
+	MOVBZ	0(R3), R3
+	CMP	R3, R3, CR7
+	BC	4, 30, 1(PC) // bne- cr7,0x4
+	ISYNC
+	MOVB	R3, ret+8(FP)
+	RET
+
+// uint64 ·Load64(uint64 volatile* ptr)
+TEXT ·Load64(SB),NOSPLIT|NOFRAME,$-8-16
+	MOVD	ptr+0(FP), R3
+	SYNC
+	MOVD	0(R3), R3
+	CMP	R3, R3, CR7
+	BC	4, 30, 1(PC) // bne- cr7,0x4
+	ISYNC
+	MOVD	R3, ret+8(FP)
+	RET
+
+// void *·Loadp(void *volatile *ptr)
+TEXT ·Loadp(SB),NOSPLIT|NOFRAME,$-8-16
+	MOVD	ptr+0(FP), R3
+	SYNC
+	MOVD	0(R3), R3
+	CMP	R3, R3, CR7
+	BC	4, 30, 1(PC) // bne- cr7,0x4
+	ISYNC
+	MOVD	R3, ret+8(FP)
+	RET
+
+// uint32 ·LoadAcq(uint32 volatile* ptr)
+TEXT ·LoadAcq(SB),NOSPLIT|NOFRAME,$-8-12
+	MOVD   ptr+0(FP), R3
+	MOVWZ  0(R3), R3
+	CMPW   R3, R3, CR7
+	BC     4, 30, 1(PC) // bne- cr7, 0x4
+	ISYNC
+	MOVW   R3, ret+8(FP)
+	RET
+
+// uint64 ·LoadAcq64(uint64 volatile* ptr)
+TEXT ·LoadAcq64(SB),NOSPLIT|NOFRAME,$-8-16
+	MOVD   ptr+0(FP), R3
+	MOVD   0(R3), R3
+	CMP    R3, R3, CR7
+	BC     4, 30, 1(PC) // bne- cr7, 0x4
+	ISYNC
+	MOVD   R3, ret+8(FP)
+	RET
+
+// bool cas(uint32 *ptr, uint32 old, uint32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOVD	ptr+0(FP), R3
+	MOVWZ	old+8(FP), R4
+	MOVWZ	new+12(FP), R5
+	LWSYNC
+cas_again:
+	LWAR	(R3), R6
+	CMPW	R6, R4
+	BNE	cas_fail
+	STWCCC	R5, (R3)
+	BNE	cas_again
+	MOVD	$1, R3
+	LWSYNC
+	MOVB	R3, ret+16(FP)
+	RET
+cas_fail:
+	MOVB	R0, ret+16(FP)
+	RET
+
+// bool	·Cas64(uint64 *ptr, uint64 old, uint64 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else {
+//		return 0;
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVD	ptr+0(FP), R3
+	MOVD	old+8(FP), R4
+	MOVD	new+16(FP), R5
+	LWSYNC
+cas64_again:
+	LDAR	(R3), R6
+	CMP	R6, R4
+	BNE	cas64_fail
+	STDCCC	R5, (R3)
+	BNE	cas64_again
+	MOVD	$1, R3
+	LWSYNC
+	MOVB	R3, ret+24(FP)
+	RET
+cas64_fail:
+	MOVB	R0, ret+24(FP)
+	RET
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	MOVD    ptr+0(FP), R3
+	MOVWZ   old+8(FP), R4
+	MOVWZ   new+12(FP), R5
+	LWSYNC
+cas_again:
+	LWAR    (R3), $0, R6        // 0 = Mutex release hint
+	CMPW    R6, R4
+	BNE     cas_fail
+	STWCCC  R5, (R3)
+	BNE     cas_again
+	MOVD    $1, R3
+	MOVB    R3, ret+16(FP)
+	RET
+cas_fail:
+	MOVB    R0, ret+16(FP)
+	RET
+
+TEXT ·Casint32(SB), NOSPLIT, $0-17
+	BR	·Cas(SB)
+
+TEXT ·Casint64(SB), NOSPLIT, $0-25
+	BR	·Cas64(SB)
+
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	BR	·Cas64(SB)
+
+TEXT ·Loaduintptr(SB),  NOSPLIT|NOFRAME, $0-16
+	BR	·Load64(SB)
+
+TEXT ·LoadAcquintptr(SB),  NOSPLIT|NOFRAME, $0-16
+	BR	·LoadAcq64(SB)
+
+TEXT ·Loaduint(SB), NOSPLIT|NOFRAME, $0-16
+	BR	·Load64(SB)
+
+TEXT ·Storeint32(SB), NOSPLIT, $0-12
+	BR	·Store(SB)
+
+TEXT ·Storeint64(SB), NOSPLIT, $0-16
+	BR	·Store64(SB)
+
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	BR	·Store64(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	BR	·StoreRel64(SB)
+
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	BR	·Xadd64(SB)
+
+TEXT ·Loadint32(SB), NOSPLIT, $0-12
+	BR	·Load(SB)
+
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	BR	·Load64(SB)
+
+TEXT ·Xaddint32(SB), NOSPLIT, $0-20
+	BR	·Xadd(SB)
+
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	BR	·Xadd64(SB)
+
+// bool casp(void **val, void *old, void *new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	} else
+//		return 0;
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	BR ·Cas64(SB)
+
+// uint32 xadd(uint32 volatile *ptr, int32 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R4
+	MOVW	delta+8(FP), R5
+	LWSYNC
+	LWAR	(R4), R3
+	ADD	R5, R3
+	STWCCC	R3, (R4)
+	BNE	-3(PC)
+	MOVW	R3, ret+16(FP)
+	RET
+
+// uint64 Xadd64(uint64 volatile *val, int64 delta)
+// Atomically:
+//	*val += delta;
+//	return *val;
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R4
+	MOVD	delta+8(FP), R5
+	LWSYNC
+	LDAR	(R4), R3
+	ADD	R5, R3
+	STDCCC	R3, (R4)
+	BNE	-3(PC)
+	MOVD	R3, ret+16(FP)
+	RET
+
+// uint32 Xchg(ptr *uint32, new uint32)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R4
+	MOVW	new+8(FP), R5
+	LWSYNC
+	LWAR	(R4), R3
+	STWCCC	R5, (R4)
+	BNE	-2(PC)
+	ISYNC
+	MOVW	R3, ret+16(FP)
+	RET
+
+// uint64 Xchg64(ptr *uint64, new uint64)
+// Atomically:
+//	old := *ptr;
+//	*ptr = new;
+//	return old;
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R4
+	MOVD	new+8(FP), R5
+	LWSYNC
+	LDAR	(R4), R3
+	STDCCC	R5, (R4)
+	BNE	-2(PC)
+	ISYNC
+	MOVD	R3, ret+16(FP)
+	RET
+
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	BR	·Xchg(SB)
+
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	BR	·Xchg64(SB)
+
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	BR	·Xchg64(SB)
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	BR	·Store64(SB)
+
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	SYNC
+	MOVW	R4, 0(R3)
+	RET
+
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R3
+	MOVB	val+8(FP), R4
+	SYNC
+	MOVB	R4, 0(R3)
+	RET
+
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOVD	ptr+0(FP), R3
+	MOVD	val+8(FP), R4
+	SYNC
+	MOVD	R4, 0(R3)
+	RET
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	LWSYNC
+	MOVW	R4, 0(R3)
+	RET
+
+TEXT ·StoreRel64(SB), NOSPLIT, $0-16
+	MOVD	ptr+0(FP), R3
+	MOVD	val+8(FP), R4
+	LWSYNC
+	MOVD	R4, 0(R3)
+	RET
+
+// void ·Or8(byte volatile*, byte);
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R3
+	MOVBZ	val+8(FP), R4
+	LWSYNC
+again:
+	LBAR	(R3), R6
+	OR	R4, R6
+	STBCCC	R6, (R3)
+	BNE	again
+	RET
+
+// void ·And8(byte volatile*, byte);
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R3
+	MOVBZ	val+8(FP), R4
+	LWSYNC
+again:
+	LBAR	(R3), R6
+	AND	R4, R6
+	STBCCC	R6, (R3)
+	BNE	again
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	LWSYNC
+again:
+	LWAR	(R3), R6
+	OR	R4, R6
+	STWCCC	R6, (R3)
+	BNE	again
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	LWSYNC
+again:
+	LWAR	(R3),R6
+	AND	R4, R6
+	STWCCC	R6, (R3)
+	BNE	again
+	RET
diff --git a/src/runtime/internal/atomic/atomic_riscv64.go b/src/runtime/internal/atomic/atomic_riscv64.go
new file mode 100644
index 0000000..8f24d61
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_riscv64.go
@@ -0,0 +1,85 @@
+// Copyright 2015 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 atomic
+
+import "unsafe"
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Load(ptr *uint32) uint32
+
+//go:noescape
+func Load8(ptr *uint8) uint8
+
+//go:noescape
+func Load64(ptr *uint64) uint64
+
+// NO go:noescape annotation; *ptr escapes if result escapes (#31525)
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+
+//go:noescape
+func LoadAcq(ptr *uint32) uint32
+
+//go:noescape
+func LoadAcq64(ptr *uint64) uint64
+
+//go:noescape
+func LoadAcquintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:noescape
+func StoreRel(ptr *uint32, val uint32)
+
+//go:noescape
+func StoreRel64(ptr *uint64, val uint64)
+
+//go:noescape
+func StoreReluintptr(ptr *uintptr, val uintptr)
diff --git a/src/runtime/internal/atomic/atomic_riscv64.s b/src/runtime/internal/atomic/atomic_riscv64.s
new file mode 100644
index 0000000..21d5adc
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_riscv64.s
@@ -0,0 +1,284 @@
+// Copyright 2014 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.
+
+// RISC-V's atomic operations have two bits, aq ("acquire") and rl ("release"),
+// which may be toggled on and off. Their precise semantics are defined in
+// section 6.3 of the specification, but the basic idea is as follows:
+//
+//   - If neither aq nor rl is set, the CPU may reorder the atomic arbitrarily.
+//     It guarantees only that it will execute atomically.
+//
+//   - If aq is set, the CPU may move the instruction backward, but not forward.
+//
+//   - If rl is set, the CPU may move the instruction forward, but not backward.
+//
+//   - If both are set, the CPU may not reorder the instruction at all.
+//
+// These four modes correspond to other well-known memory models on other CPUs.
+// On ARM, aq corresponds to a dmb ishst, aq+rl corresponds to a dmb ish. On
+// Intel, aq corresponds to an lfence, rl to an sfence, and aq+rl to an mfence
+// (or a lock prefix).
+//
+// Go's memory model requires that
+//   - if a read happens after a write, the read must observe the write, and
+//     that
+//   - if a read happens concurrently with a write, the read may observe the
+//     write.
+// aq is sufficient to guarantee this, so that's what we use here. (This jibes
+// with ARM, which uses dmb ishst.)
+
+#include "textflag.h"
+
+// func Cas(ptr *uint64, old, new uint64) bool
+// Atomically:
+//      if(*val == old){
+//              *val = new;
+//              return 1;
+//      } else {
+//              return 0;
+//      }
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOV	ptr+0(FP), A0
+	MOVW	old+8(FP), A1
+	MOVW	new+12(FP), A2
+cas_again:
+	LRW	(A0), A3
+	BNE	A3, A1, cas_fail
+	SCW	A2, (A0), A4
+	BNE	A4, ZERO, cas_again
+	MOV	$1, A0
+	MOVB	A0, ret+16(FP)
+	RET
+cas_fail:
+	MOV	$0, A0
+	MOV	A0, ret+16(FP)
+	RET
+
+// func Cas64(ptr *uint64, old, new uint64) bool
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOV	ptr+0(FP), A0
+	MOV	old+8(FP), A1
+	MOV	new+16(FP), A2
+cas_again:
+	LRD	(A0), A3
+	BNE	A3, A1, cas_fail
+	SCD	A2, (A0), A4
+	BNE	A4, ZERO, cas_again
+	MOV	$1, A0
+	MOVB	A0, ret+24(FP)
+	RET
+cas_fail:
+	MOVB	ZERO, ret+24(FP)
+	RET
+
+// func Load(ptr *uint32) uint32
+TEXT ·Load(SB),NOSPLIT|NOFRAME,$0-12
+	MOV	ptr+0(FP), A0
+	LRW	(A0), A0
+	MOVW	A0, ret+8(FP)
+	RET
+
+// func Load8(ptr *uint8) uint8
+TEXT ·Load8(SB),NOSPLIT|NOFRAME,$0-9
+	MOV	ptr+0(FP), A0
+	FENCE
+	MOVBU	(A0), A1
+	FENCE
+	MOVB	A1, ret+8(FP)
+	RET
+
+// func Load64(ptr *uint64) uint64
+TEXT ·Load64(SB),NOSPLIT|NOFRAME,$0-16
+	MOV	ptr+0(FP), A0
+	LRD	(A0), A0
+	MOV	A0, ret+8(FP)
+	RET
+
+// func Store(ptr *uint32, val uint32)
+TEXT ·Store(SB), NOSPLIT, $0-12
+	MOV	ptr+0(FP), A0
+	MOVW	val+8(FP), A1
+	AMOSWAPW A1, (A0), ZERO
+	RET
+
+// func Store8(ptr *uint8, val uint8)
+TEXT ·Store8(SB), NOSPLIT, $0-9
+	MOV	ptr+0(FP), A0
+	MOVBU	val+8(FP), A1
+	FENCE
+	MOVB	A1, (A0)
+	FENCE
+	RET
+
+// func Store64(ptr *uint64, val uint64)
+TEXT ·Store64(SB), NOSPLIT, $0-16
+	MOV	ptr+0(FP), A0
+	MOV	val+8(FP), A1
+	AMOSWAPD A1, (A0), ZERO
+	RET
+
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	JMP	·Cas64(SB)
+
+TEXT ·Casint32(SB),NOSPLIT,$0-17
+	JMP	·Cas(SB)
+
+TEXT ·Casint64(SB),NOSPLIT,$0-25
+	JMP	·Cas64(SB)
+
+TEXT ·Casuintptr(SB),NOSPLIT,$0-25
+	JMP	·Cas64(SB)
+
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	JMP	·Cas(SB)
+
+TEXT ·Loaduintptr(SB),NOSPLIT,$0-16
+	JMP	·Load64(SB)
+
+TEXT ·Storeint32(SB),NOSPLIT,$0-12
+	JMP	·Store(SB)
+
+TEXT ·Storeint64(SB),NOSPLIT,$0-16
+	JMP	·Store64(SB)
+
+TEXT ·Storeuintptr(SB),NOSPLIT,$0-16
+	JMP	·Store64(SB)
+
+TEXT ·Loaduint(SB),NOSPLIT,$0-16
+	JMP ·Loaduintptr(SB)
+
+TEXT ·Loadint32(SB),NOSPLIT,$0-12
+	JMP ·Load(SB)
+
+TEXT ·Loadint64(SB),NOSPLIT,$0-16
+	JMP ·Load64(SB)
+
+TEXT ·Xaddint32(SB),NOSPLIT,$0-20
+	JMP ·Xadd(SB)
+
+TEXT ·Xaddint64(SB),NOSPLIT,$0-24
+	MOV	ptr+0(FP), A0
+	MOV	delta+8(FP), A1
+	AMOADDD A1, (A0), A0
+	ADD	A0, A1, A0
+	MOVW	A0, ret+16(FP)
+	RET
+
+TEXT ·LoadAcq(SB),NOSPLIT|NOFRAME,$0-12
+	JMP	·Load(SB)
+
+TEXT ·LoadAcq64(SB),NOSPLIT|NOFRAME,$0-16
+	JMP	·Load64(SB)
+
+TEXT ·LoadAcquintptr(SB),NOSPLIT|NOFRAME,$0-16
+	JMP	·Load64(SB)
+
+// func Loadp(ptr unsafe.Pointer) unsafe.Pointer
+TEXT ·Loadp(SB),NOSPLIT,$0-16
+	JMP	·Load64(SB)
+
+// func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreRel(SB), NOSPLIT, $0-12
+	JMP	·Store(SB)
+
+TEXT ·StoreRel64(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
+	JMP	·Store64(SB)
+
+// func Xchg(ptr *uint32, new uint32) uint32
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOV	ptr+0(FP), A0
+	MOVW	new+8(FP), A1
+	AMOSWAPW A1, (A0), A1
+	MOVW	A1, ret+16(FP)
+	RET
+
+// func Xchg64(ptr *uint64, new uint64) uint64
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOV	ptr+0(FP), A0
+	MOV	new+8(FP), A1
+	AMOSWAPD A1, (A0), A1
+	MOV	A1, ret+16(FP)
+	RET
+
+// Atomically:
+//      *val += delta;
+//      return *val;
+
+// func Xadd(ptr *uint32, delta int32) uint32
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOV	ptr+0(FP), A0
+	MOVW	delta+8(FP), A1
+	AMOADDW A1, (A0), A2
+	ADD	A2,A1,A0
+	MOVW	A0, ret+16(FP)
+	RET
+
+// func Xadd64(ptr *uint64, delta int64) uint64
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOV	ptr+0(FP), A0
+	MOV	delta+8(FP), A1
+	AMOADDD A1, (A0), A2
+	ADD	A2, A1, A0
+	MOV	A0, ret+16(FP)
+	RET
+
+// func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	JMP	·Xadd64(SB)
+
+// func Xchgint32(ptr *int32, new int32) int32
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	JMP	·Xchg(SB)
+
+// func Xchgint64(ptr *int64, new int64) int64
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+// func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	JMP	·Xchg64(SB)
+
+// func And8(ptr *uint8, val uint8)
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOV	ptr+0(FP), A0
+	MOVBU	val+8(FP), A1
+	AND	$3, A0, A2
+	AND	$-4, A0
+	SLL	$3, A2
+	XOR	$255, A1
+	SLL	A2, A1
+	XOR	$-1, A1
+	AMOANDW A1, (A0), ZERO
+	RET
+
+// func Or8(ptr *uint8, val uint8)
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOV	ptr+0(FP), A0
+	MOVBU	val+8(FP), A1
+	AND	$3, A0, A2
+	AND	$-4, A0
+	SLL	$3, A2
+	SLL	A2, A1
+	AMOORW	A1, (A0), ZERO
+	RET
+
+// func And(ptr *uint32, val uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOV	ptr+0(FP), A0
+	MOVW	val+8(FP), A1
+	AMOANDW	A1, (A0), ZERO
+	RET
+
+// func Or(ptr *uint32, val uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOV	ptr+0(FP), A0
+	MOVW	val+8(FP), A1
+	AMOORW	A1, (A0), ZERO
+	RET
diff --git a/src/runtime/internal/atomic/atomic_s390x.go b/src/runtime/internal/atomic/atomic_s390x.go
new file mode 100644
index 0000000..9855bf0
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_s390x.go
@@ -0,0 +1,123 @@
+// Copyright 2016 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 atomic
+
+import "unsafe"
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Load
+//go:linkname Loadp
+//go:linkname Load64
+
+//go:nosplit
+//go:noinline
+func Load(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer {
+	return *(*unsafe.Pointer)(ptr)
+}
+
+//go:nosplit
+//go:noinline
+func Load8(ptr *uint8) uint8 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Load64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcquintptr(ptr *uintptr) uintptr {
+	return *ptr
+}
+
+//go:noescape
+func Store(ptr *uint32, val uint32)
+
+//go:noescape
+func Store8(ptr *uint8, val uint8)
+
+//go:noescape
+func Store64(ptr *uint64, val uint64)
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:nosplit
+//go:noinline
+func StoreRel(ptr *uint32, val uint32) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func StoreRel64(ptr *uint64, val uint64) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func StoreReluintptr(ptr *uintptr, val uintptr) {
+	*ptr = val
+}
+
+//go:noescape
+func And8(ptr *uint8, val uint8)
+
+//go:noescape
+func Or8(ptr *uint8, val uint8)
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:noescape
+func And(ptr *uint32, val uint32)
+
+//go:noescape
+func Or(ptr *uint32, val uint32)
+
+//go:noescape
+func Xadd(ptr *uint32, delta int32) uint32
+
+//go:noescape
+func Xadd64(ptr *uint64, delta int64) uint64
+
+//go:noescape
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+
+//go:noescape
+func Xchg(ptr *uint32, new uint32) uint32
+
+//go:noescape
+func Xchg64(ptr *uint64, new uint64) uint64
+
+//go:noescape
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+
+//go:noescape
+func Cas64(ptr *uint64, old, new uint64) bool
+
+//go:noescape
+func CasRel(ptr *uint32, old, new uint32) bool
diff --git a/src/runtime/internal/atomic/atomic_s390x.s b/src/runtime/internal/atomic/atomic_s390x.s
new file mode 100644
index 0000000..a0c204b
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_s390x.s
@@ -0,0 +1,248 @@
+// Copyright 2016 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.
+
+#include "textflag.h"
+
+// func Store(ptr *uint32, val uint32)
+TEXT ·Store(SB), NOSPLIT, $0
+	MOVD	ptr+0(FP), R2
+	MOVWZ	val+8(FP), R3
+	MOVW	R3, 0(R2)
+	SYNC
+	RET
+
+// func Store8(ptr *uint8, val uint8)
+TEXT ·Store8(SB), NOSPLIT, $0
+	MOVD	ptr+0(FP), R2
+	MOVB	val+8(FP), R3
+	MOVB	R3, 0(R2)
+	SYNC
+	RET
+
+// func Store64(ptr *uint64, val uint64)
+TEXT ·Store64(SB), NOSPLIT, $0
+	MOVD	ptr+0(FP), R2
+	MOVD	val+8(FP), R3
+	MOVD	R3, 0(R2)
+	SYNC
+	RET
+
+// func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+TEXT ·StorepNoWB(SB), NOSPLIT, $0
+	MOVD	ptr+0(FP), R2
+	MOVD	val+8(FP), R3
+	MOVD	R3, 0(R2)
+	SYNC
+	RET
+
+// func Cas(ptr *uint32, old, new uint32) bool
+// Atomically:
+//	if *ptr == old {
+//		*val = new
+//		return 1
+//	} else {
+//		return 0
+//	}
+TEXT ·Cas(SB), NOSPLIT, $0-17
+	MOVD	ptr+0(FP), R3
+	MOVWZ	old+8(FP), R4
+	MOVWZ	new+12(FP), R5
+	CS	R4, R5, 0(R3)    //  if (R4 == 0(R3)) then 0(R3)= R5
+	BNE	cas_fail
+	MOVB	$1, ret+16(FP)
+	RET
+cas_fail:
+	MOVB	$0, ret+16(FP)
+	RET
+
+// func Cas64(ptr *uint64, old, new uint64) bool
+// Atomically:
+//	if *ptr == old {
+//		*ptr = new
+//		return 1
+//	} else {
+//		return 0
+//	}
+TEXT ·Cas64(SB), NOSPLIT, $0-25
+	MOVD	ptr+0(FP), R3
+	MOVD	old+8(FP), R4
+	MOVD	new+16(FP), R5
+	CSG	R4, R5, 0(R3)    //  if (R4 == 0(R3)) then 0(R3)= R5
+	BNE	cas64_fail
+	MOVB	$1, ret+24(FP)
+	RET
+cas64_fail:
+	MOVB	$0, ret+24(FP)
+	RET
+
+// func Casint32(ptr *int32, old, new int32) bool
+TEXT ·Casint32(SB), NOSPLIT, $0-17
+	BR	·Cas(SB)
+
+// func Casint64(ptr *int64, old, new int64) bool
+TEXT ·Casint64(SB), NOSPLIT, $0-25
+	BR	·Cas64(SB)
+
+// func Casuintptr(ptr *uintptr, old, new uintptr) bool
+TEXT ·Casuintptr(SB), NOSPLIT, $0-25
+	BR	·Cas64(SB)
+
+// func CasRel(ptr *uint32, old, new uint32) bool
+TEXT ·CasRel(SB), NOSPLIT, $0-17
+	BR	·Cas(SB)
+
+// func Loaduintptr(ptr *uintptr) uintptr
+TEXT ·Loaduintptr(SB), NOSPLIT, $0-16
+	BR	·Load64(SB)
+
+// func Loaduint(ptr *uint) uint
+TEXT ·Loaduint(SB), NOSPLIT, $0-16
+	BR	·Load64(SB)
+
+// func Storeint32(ptr *int32, new int32)
+TEXT ·Storeint32(SB), NOSPLIT, $0-12
+	BR	·Store(SB)
+
+// func Storeint64(ptr *int64, new int64)
+TEXT ·Storeint64(SB), NOSPLIT, $0-16
+	BR	·Store64(SB)
+
+// func Storeuintptr(ptr *uintptr, new uintptr)
+TEXT ·Storeuintptr(SB), NOSPLIT, $0-16
+	BR	·Store64(SB)
+
+// func Loadint32(ptr *int32) int32
+TEXT ·Loadint32(SB), NOSPLIT, $0-12
+	BR	·Load(SB)
+
+// func Loadint64(ptr *int64) int64
+TEXT ·Loadint64(SB), NOSPLIT, $0-16
+	BR	·Load64(SB)
+
+// func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
+TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
+	BR	·Xadd64(SB)
+
+// func Xaddint32(ptr *int32, delta int32) int32
+TEXT ·Xaddint32(SB), NOSPLIT, $0-20
+	BR	·Xadd(SB)
+
+// func Xaddint64(ptr *int64, delta int64) int64
+TEXT ·Xaddint64(SB), NOSPLIT, $0-24
+	BR	·Xadd64(SB)
+
+// func Casp1(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool
+// Atomically:
+//	if *ptr == old {
+//		*ptr = new
+//		return 1
+//	} else {
+//		return 0
+//	}
+TEXT ·Casp1(SB), NOSPLIT, $0-25
+	BR ·Cas64(SB)
+
+// func Xadd(ptr *uint32, delta int32) uint32
+// Atomically:
+//	*ptr += delta
+//	return *ptr
+TEXT ·Xadd(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R4
+	MOVW	delta+8(FP), R5
+	MOVW	(R4), R3
+repeat:
+	ADD	R5, R3, R6
+	CS	R3, R6, (R4) // if R3==(R4) then (R4)=R6 else R3=(R4)
+	BNE	repeat
+	MOVW	R6, ret+16(FP)
+	RET
+
+// func Xadd64(ptr *uint64, delta int64) uint64
+TEXT ·Xadd64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R4
+	MOVD	delta+8(FP), R5
+	MOVD	(R4), R3
+repeat:
+	ADD	R5, R3, R6
+	CSG	R3, R6, (R4) // if R3==(R4) then (R4)=R6 else R3=(R4)
+	BNE	repeat
+	MOVD	R6, ret+16(FP)
+	RET
+
+// func Xchg(ptr *uint32, new uint32) uint32
+TEXT ·Xchg(SB), NOSPLIT, $0-20
+	MOVD	ptr+0(FP), R4
+	MOVW	new+8(FP), R3
+	MOVW	(R4), R6
+repeat:
+	CS	R6, R3, (R4) // if R6==(R4) then (R4)=R3 else R6=(R4)
+	BNE	repeat
+	MOVW	R6, ret+16(FP)
+	RET
+
+// func Xchg64(ptr *uint64, new uint64) uint64
+TEXT ·Xchg64(SB), NOSPLIT, $0-24
+	MOVD	ptr+0(FP), R4
+	MOVD	new+8(FP), R3
+	MOVD	(R4), R6
+repeat:
+	CSG	R6, R3, (R4) // if R6==(R4) then (R4)=R3 else R6=(R4)
+	BNE	repeat
+	MOVD	R6, ret+16(FP)
+	RET
+
+// func Xchgint32(ptr *int32, new int32) int32
+TEXT ·Xchgint32(SB), NOSPLIT, $0-20
+	BR	·Xchg(SB)
+
+// func Xchgint64(ptr *int64, new int64) int64
+TEXT ·Xchgint64(SB), NOSPLIT, $0-24
+	BR	·Xchg64(SB)
+
+// func Xchguintptr(ptr *uintptr, new uintptr) uintptr
+TEXT ·Xchguintptr(SB), NOSPLIT, $0-24
+	BR	·Xchg64(SB)
+
+// func Or8(addr *uint8, v uint8)
+TEXT ·Or8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R3
+	MOVBZ	val+8(FP), R4
+	// We don't have atomic operations that work on individual bytes so we
+	// need to align addr down to a word boundary and create a mask
+	// containing v to OR with the entire word atomically.
+	MOVD	$(3<<3), R5
+	RXSBG	$59, $60, $3, R3, R5 // R5 = 24 - ((addr % 4) * 8) = ((addr & 3) << 3) ^ (3 << 3)
+	ANDW	$~3, R3              // R3 = floor(addr, 4) = addr &^ 3
+	SLW	R5, R4               // R4 = uint32(v) << R5
+	LAO	R4, R6, 0(R3)        // R6 = *R3; *R3 |= R4; (atomic)
+	RET
+
+// func And8(addr *uint8, v uint8)
+TEXT ·And8(SB), NOSPLIT, $0-9
+	MOVD	ptr+0(FP), R3
+	MOVBZ	val+8(FP), R4
+	// We don't have atomic operations that work on individual bytes so we
+	// need to align addr down to a word boundary and create a mask
+	// containing v to AND with the entire word atomically.
+	ORW	$~0xff, R4           // R4 = uint32(v) | 0xffffff00
+	MOVD	$(3<<3), R5
+	RXSBG	$59, $60, $3, R3, R5 // R5 = 24 - ((addr % 4) * 8) = ((addr & 3) << 3) ^ (3 << 3)
+	ANDW	$~3, R3              // R3 = floor(addr, 4) = addr &^ 3
+	RLL	R5, R4, R4           // R4 = rotl(R4, R5)
+	LAN	R4, R6, 0(R3)        // R6 = *R3; *R3 &= R4; (atomic)
+	RET
+
+// func Or(addr *uint32, v uint32)
+TEXT ·Or(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	LAO	R4, R6, 0(R3)        // R6 = *R3; *R3 |= R4; (atomic)
+	RET
+
+// func And(addr *uint32, v uint32)
+TEXT ·And(SB), NOSPLIT, $0-12
+	MOVD	ptr+0(FP), R3
+	MOVW	val+8(FP), R4
+	LAN	R4, R6, 0(R3)        // R6 = *R3; *R3 &= R4; (atomic)
+	RET
diff --git a/src/runtime/internal/atomic/atomic_test.go b/src/runtime/internal/atomic/atomic_test.go
new file mode 100644
index 0000000..2427bfd
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_test.go
@@ -0,0 +1,386 @@
+// Copyright 2015 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 atomic_test
+
+import (
+	"internal/goarch"
+	"runtime"
+	"runtime/internal/atomic"
+	"testing"
+	"unsafe"
+)
+
+func runParallel(N, iter int, f func()) {
+	defer runtime.GOMAXPROCS(runtime.GOMAXPROCS(int(N)))
+	done := make(chan bool)
+	for i := 0; i < N; i++ {
+		go func() {
+			for j := 0; j < iter; j++ {
+				f()
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < N; i++ {
+		<-done
+	}
+}
+
+func TestXadduintptr(t *testing.T) {
+	N := 20
+	iter := 100000
+	if testing.Short() {
+		N = 10
+		iter = 10000
+	}
+	inc := uintptr(100)
+	total := uintptr(0)
+	runParallel(N, iter, func() {
+		atomic.Xadduintptr(&total, inc)
+	})
+	if want := uintptr(N*iter) * inc; want != total {
+		t.Fatalf("xadduintpr error, want %d, got %d", want, total)
+	}
+	total = 0
+	runParallel(N, iter, func() {
+		atomic.Xadduintptr(&total, inc)
+		atomic.Xadduintptr(&total, uintptr(-int64(inc)))
+	})
+	if total != 0 {
+		t.Fatalf("xadduintpr total error, want %d, got %d", 0, total)
+	}
+}
+
+// Tests that xadduintptr correctly updates 64-bit values. The place where
+// we actually do so is mstats.go, functions mSysStat{Inc,Dec}.
+func TestXadduintptrOnUint64(t *testing.T) {
+	if goarch.BigEndian {
+		// On big endian architectures, we never use xadduintptr to update
+		// 64-bit values and hence we skip the test.  (Note that functions
+		// mSysStat{Inc,Dec} in mstats.go have explicit checks for
+		// big-endianness.)
+		t.Skip("skip xadduintptr on big endian architecture")
+	}
+	const inc = 100
+	val := uint64(0)
+	atomic.Xadduintptr((*uintptr)(unsafe.Pointer(&val)), inc)
+	if inc != val {
+		t.Fatalf("xadduintptr should increase lower-order bits, want %d, got %d", inc, val)
+	}
+}
+
+func shouldPanic(t *testing.T, name string, f func()) {
+	defer func() {
+		// Check that all GC maps are sane.
+		runtime.GC()
+
+		err := recover()
+		want := "unaligned 64-bit atomic operation"
+		if err == nil {
+			t.Errorf("%s did not panic", name)
+		} else if s, _ := err.(string); s != want {
+			t.Errorf("%s: wanted panic %q, got %q", name, want, err)
+		}
+	}()
+	f()
+}
+
+// Variant of sync/atomic's TestUnaligned64:
+func TestUnaligned64(t *testing.T) {
+	// Unaligned 64-bit atomics on 32-bit systems are
+	// a continual source of pain. Test that on 32-bit systems they crash
+	// instead of failing silently.
+
+	if unsafe.Sizeof(int(0)) != 4 {
+		t.Skip("test only runs on 32-bit systems")
+	}
+
+	x := make([]uint32, 4)
+	u := unsafe.Pointer(uintptr(unsafe.Pointer(&x[0])) | 4) // force alignment to 4
+
+	up64 := (*uint64)(u) // misaligned
+	p64 := (*int64)(u)   // misaligned
+
+	shouldPanic(t, "Load64", func() { atomic.Load64(up64) })
+	shouldPanic(t, "Loadint64", func() { atomic.Loadint64(p64) })
+	shouldPanic(t, "Store64", func() { atomic.Store64(up64, 0) })
+	shouldPanic(t, "Xadd64", func() { atomic.Xadd64(up64, 1) })
+	shouldPanic(t, "Xchg64", func() { atomic.Xchg64(up64, 1) })
+	shouldPanic(t, "Cas64", func() { atomic.Cas64(up64, 1, 2) })
+}
+
+func TestAnd8(t *testing.T) {
+	// Basic sanity check.
+	x := uint8(0xff)
+	for i := uint8(0); i < 8; i++ {
+		atomic.And8(&x, ^(1 << i))
+		if r := uint8(0xff) << (i + 1); x != r {
+			t.Fatalf("clearing bit %#x: want %#x, got %#x", uint8(1<<i), r, x)
+		}
+	}
+
+	// Set every bit in array to 1.
+	a := make([]uint8, 1<<12)
+	for i := range a {
+		a[i] = 0xff
+	}
+
+	// Clear array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 8; i++ {
+		m := ^uint8(1 << i)
+		go func() {
+			for i := range a {
+				atomic.And8(&a[i], m)
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 8; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally cleared.
+	for i, v := range a {
+		if v != 0 {
+			t.Fatalf("a[%v] not cleared: want %#x, got %#x", i, uint8(0), v)
+		}
+	}
+}
+
+func TestAnd(t *testing.T) {
+	// Basic sanity check.
+	x := uint32(0xffffffff)
+	for i := uint32(0); i < 32; i++ {
+		atomic.And(&x, ^(1 << i))
+		if r := uint32(0xffffffff) << (i + 1); x != r {
+			t.Fatalf("clearing bit %#x: want %#x, got %#x", uint32(1<<i), r, x)
+		}
+	}
+
+	// Set every bit in array to 1.
+	a := make([]uint32, 1<<12)
+	for i := range a {
+		a[i] = 0xffffffff
+	}
+
+	// Clear array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 32; i++ {
+		m := ^uint32(1 << i)
+		go func() {
+			for i := range a {
+				atomic.And(&a[i], m)
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 32; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally cleared.
+	for i, v := range a {
+		if v != 0 {
+			t.Fatalf("a[%v] not cleared: want %#x, got %#x", i, uint32(0), v)
+		}
+	}
+}
+
+func TestOr8(t *testing.T) {
+	// Basic sanity check.
+	x := uint8(0)
+	for i := uint8(0); i < 8; i++ {
+		atomic.Or8(&x, 1<<i)
+		if r := (uint8(1) << (i + 1)) - 1; x != r {
+			t.Fatalf("setting bit %#x: want %#x, got %#x", uint8(1)<<i, r, x)
+		}
+	}
+
+	// Start with every bit in array set to 0.
+	a := make([]uint8, 1<<12)
+
+	// Set every bit in array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 8; i++ {
+		m := uint8(1 << i)
+		go func() {
+			for i := range a {
+				atomic.Or8(&a[i], m)
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 8; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally set.
+	for i, v := range a {
+		if v != 0xff {
+			t.Fatalf("a[%v] not fully set: want %#x, got %#x", i, uint8(0xff), v)
+		}
+	}
+}
+
+func TestOr(t *testing.T) {
+	// Basic sanity check.
+	x := uint32(0)
+	for i := uint32(0); i < 32; i++ {
+		atomic.Or(&x, 1<<i)
+		if r := (uint32(1) << (i + 1)) - 1; x != r {
+			t.Fatalf("setting bit %#x: want %#x, got %#x", uint32(1)<<i, r, x)
+		}
+	}
+
+	// Start with every bit in array set to 0.
+	a := make([]uint32, 1<<12)
+
+	// Set every bit in array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 32; i++ {
+		m := uint32(1 << i)
+		go func() {
+			for i := range a {
+				atomic.Or(&a[i], m)
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 32; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally set.
+	for i, v := range a {
+		if v != 0xffffffff {
+			t.Fatalf("a[%v] not fully set: want %#x, got %#x", i, uint32(0xffffffff), v)
+		}
+	}
+}
+
+func TestBitwiseContended8(t *testing.T) {
+	// Start with every bit in array set to 0.
+	a := make([]uint8, 16)
+
+	// Iterations to try.
+	N := 1 << 16
+	if testing.Short() {
+		N = 1 << 10
+	}
+
+	// Set and then clear every bit in the array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 8; i++ {
+		m := uint8(1 << i)
+		go func() {
+			for n := 0; n < N; n++ {
+				for i := range a {
+					atomic.Or8(&a[i], m)
+					if atomic.Load8(&a[i])&m != m {
+						t.Errorf("a[%v] bit %#x not set", i, m)
+					}
+					atomic.And8(&a[i], ^m)
+					if atomic.Load8(&a[i])&m != 0 {
+						t.Errorf("a[%v] bit %#x not clear", i, m)
+					}
+				}
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 8; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally cleared.
+	for i, v := range a {
+		if v != 0 {
+			t.Fatalf("a[%v] not cleared: want %#x, got %#x", i, uint8(0), v)
+		}
+	}
+}
+
+func TestBitwiseContended(t *testing.T) {
+	// Start with every bit in array set to 0.
+	a := make([]uint32, 16)
+
+	// Iterations to try.
+	N := 1 << 16
+	if testing.Short() {
+		N = 1 << 10
+	}
+
+	// Set and then clear every bit in the array bit-by-bit in different goroutines.
+	done := make(chan bool)
+	for i := 0; i < 32; i++ {
+		m := uint32(1 << i)
+		go func() {
+			for n := 0; n < N; n++ {
+				for i := range a {
+					atomic.Or(&a[i], m)
+					if atomic.Load(&a[i])&m != m {
+						t.Errorf("a[%v] bit %#x not set", i, m)
+					}
+					atomic.And(&a[i], ^m)
+					if atomic.Load(&a[i])&m != 0 {
+						t.Errorf("a[%v] bit %#x not clear", i, m)
+					}
+				}
+			}
+			done <- true
+		}()
+	}
+	for i := 0; i < 32; i++ {
+		<-done
+	}
+
+	// Check that the array has been totally cleared.
+	for i, v := range a {
+		if v != 0 {
+			t.Fatalf("a[%v] not cleared: want %#x, got %#x", i, uint32(0), v)
+		}
+	}
+}
+
+func TestCasRel(t *testing.T) {
+	const _magic = 0x5a5aa5a5
+	var x struct {
+		before uint32
+		i      uint32
+		after  uint32
+		o      uint32
+		n      uint32
+	}
+
+	x.before = _magic
+	x.after = _magic
+	for j := 0; j < 32; j += 1 {
+		x.i = (1 << j) + 0
+		x.o = (1 << j) + 0
+		x.n = (1 << j) + 1
+		if !atomic.CasRel(&x.i, x.o, x.n) {
+			t.Fatalf("should have swapped %#x %#x", x.o, x.n)
+		}
+
+		if x.i != x.n {
+			t.Fatalf("wrong x.i after swap: x.i=%#x x.n=%#x", x.i, x.n)
+		}
+
+		if x.before != _magic || x.after != _magic {
+			t.Fatalf("wrong magic: %#x _ %#x != %#x _ %#x", x.before, x.after, _magic, _magic)
+		}
+	}
+}
+
+func TestStorepNoWB(t *testing.T) {
+	var p [2]*int
+	for i := range p {
+		atomic.StorepNoWB(unsafe.Pointer(&p[i]), unsafe.Pointer(new(int)))
+	}
+	if p[0] == p[1] {
+		t.Error("Bad escape analysis of StorepNoWB")
+	}
+}
diff --git a/src/runtime/internal/atomic/atomic_wasm.go b/src/runtime/internal/atomic/atomic_wasm.go
new file mode 100644
index 0000000..835fc43
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_wasm.go
@@ -0,0 +1,341 @@
+// 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.
+
+// TODO(neelance): implement with actual atomic operations as soon as threads are available
+// See https://github.com/WebAssembly/design/issues/1073
+
+// Export some functions via linkname to assembly in sync/atomic.
+//
+//go:linkname Load
+//go:linkname Loadp
+//go:linkname Load64
+//go:linkname Loadint32
+//go:linkname Loadint64
+//go:linkname Loaduintptr
+//go:linkname Xadd
+//go:linkname Xaddint32
+//go:linkname Xaddint64
+//go:linkname Xadd64
+//go:linkname Xadduintptr
+//go:linkname Xchg
+//go:linkname Xchg64
+//go:linkname Xchgint32
+//go:linkname Xchgint64
+//go:linkname Xchguintptr
+//go:linkname Cas
+//go:linkname Cas64
+//go:linkname Casint32
+//go:linkname Casint64
+//go:linkname Casuintptr
+//go:linkname Store
+//go:linkname Store64
+//go:linkname Storeint32
+//go:linkname Storeint64
+//go:linkname Storeuintptr
+
+package atomic
+
+import "unsafe"
+
+//go:nosplit
+//go:noinline
+func Load(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadp(ptr unsafe.Pointer) unsafe.Pointer {
+	return *(*unsafe.Pointer)(ptr)
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq(ptr *uint32) uint32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcq64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func LoadAcquintptr(ptr *uintptr) uintptr {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Load8(ptr *uint8) uint8 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Load64(ptr *uint64) uint64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Xadd(ptr *uint32, delta int32) uint32 {
+	new := *ptr + uint32(delta)
+	*ptr = new
+	return new
+}
+
+//go:nosplit
+//go:noinline
+func Xadd64(ptr *uint64, delta int64) uint64 {
+	new := *ptr + uint64(delta)
+	*ptr = new
+	return new
+}
+
+//go:nosplit
+//go:noinline
+func Xadduintptr(ptr *uintptr, delta uintptr) uintptr {
+	new := *ptr + delta
+	*ptr = new
+	return new
+}
+
+//go:nosplit
+//go:noinline
+func Xchg(ptr *uint32, new uint32) uint32 {
+	old := *ptr
+	*ptr = new
+	return old
+}
+
+//go:nosplit
+//go:noinline
+func Xchg64(ptr *uint64, new uint64) uint64 {
+	old := *ptr
+	*ptr = new
+	return old
+}
+
+//go:nosplit
+//go:noinline
+func Xchgint32(ptr *int32, new int32) int32 {
+	old := *ptr
+	*ptr = new
+	return old
+}
+
+//go:nosplit
+//go:noinline
+func Xchgint64(ptr *int64, new int64) int64 {
+	old := *ptr
+	*ptr = new
+	return old
+}
+
+//go:nosplit
+//go:noinline
+func Xchguintptr(ptr *uintptr, new uintptr) uintptr {
+	old := *ptr
+	*ptr = new
+	return old
+}
+
+//go:nosplit
+//go:noinline
+func And8(ptr *uint8, val uint8) {
+	*ptr = *ptr & val
+}
+
+//go:nosplit
+//go:noinline
+func Or8(ptr *uint8, val uint8) {
+	*ptr = *ptr | val
+}
+
+// NOTE: Do not add atomicxor8 (XOR is not idempotent).
+
+//go:nosplit
+//go:noinline
+func And(ptr *uint32, val uint32) {
+	*ptr = *ptr & val
+}
+
+//go:nosplit
+//go:noinline
+func Or(ptr *uint32, val uint32) {
+	*ptr = *ptr | val
+}
+
+//go:nosplit
+//go:noinline
+func Cas64(ptr *uint64, old, new uint64) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Store(ptr *uint32, val uint32) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func StoreRel(ptr *uint32, val uint32) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func StoreRel64(ptr *uint64, val uint64) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func StoreReluintptr(ptr *uintptr, val uintptr) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func Store8(ptr *uint8, val uint8) {
+	*ptr = val
+}
+
+//go:nosplit
+//go:noinline
+func Store64(ptr *uint64, val uint64) {
+	*ptr = val
+}
+
+// StorepNoWB performs *ptr = val atomically and without a write
+// barrier.
+//
+// NO go:noescape annotation; see atomic_pointer.go.
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
+
+//go:nosplit
+//go:noinline
+func Casint32(ptr *int32, old, new int32) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Casint64(ptr *int64, old, new int64) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Cas(ptr *uint32, old, new uint32) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Casp1(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Casuintptr(ptr *uintptr, old, new uintptr) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func CasRel(ptr *uint32, old, new uint32) bool {
+	if *ptr == old {
+		*ptr = new
+		return true
+	}
+	return false
+}
+
+//go:nosplit
+//go:noinline
+func Storeint32(ptr *int32, new int32) {
+	*ptr = new
+}
+
+//go:nosplit
+//go:noinline
+func Storeint64(ptr *int64, new int64) {
+	*ptr = new
+}
+
+//go:nosplit
+//go:noinline
+func Storeuintptr(ptr *uintptr, new uintptr) {
+	*ptr = new
+}
+
+//go:nosplit
+//go:noinline
+func Loaduintptr(ptr *uintptr) uintptr {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loaduint(ptr *uint) uint {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadint32(ptr *int32) int32 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Loadint64(ptr *int64) int64 {
+	return *ptr
+}
+
+//go:nosplit
+//go:noinline
+func Xaddint32(ptr *int32, delta int32) int32 {
+	new := *ptr + delta
+	*ptr = new
+	return new
+}
+
+//go:nosplit
+//go:noinline
+func Xaddint64(ptr *int64, delta int64) int64 {
+	new := *ptr + delta
+	*ptr = new
+	return new
+}
diff --git a/src/runtime/internal/atomic/atomic_wasm.s b/src/runtime/internal/atomic/atomic_wasm.s
new file mode 100644
index 0000000..1c2d1ce
--- /dev/null
+++ b/src/runtime/internal/atomic/atomic_wasm.s
@@ -0,0 +1,10 @@
+// Copyright 2020 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.
+
+#include "textflag.h"
+
+TEXT ·StorepNoWB(SB), NOSPLIT, $0-16
+	MOVD ptr+0(FP), R0
+	MOVD val+8(FP), 0(R0)
+	RET
diff --git a/src/runtime/internal/atomic/bench_test.go b/src/runtime/internal/atomic/bench_test.go
new file mode 100644
index 0000000..efc0531
--- /dev/null
+++ b/src/runtime/internal/atomic/bench_test.go
@@ -0,0 +1,195 @@
+// Copyright 2016 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 atomic_test
+
+import (
+	"runtime/internal/atomic"
+	"testing"
+)
+
+var sink any
+
+func BenchmarkAtomicLoad64(b *testing.B) {
+	var x uint64
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		_ = atomic.Load64(&x)
+	}
+}
+
+func BenchmarkAtomicStore64(b *testing.B) {
+	var x uint64
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.Store64(&x, 0)
+	}
+}
+
+func BenchmarkAtomicLoad(b *testing.B) {
+	var x uint32
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		_ = atomic.Load(&x)
+	}
+}
+
+func BenchmarkAtomicStore(b *testing.B) {
+	var x uint32
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.Store(&x, 0)
+	}
+}
+
+func BenchmarkAnd8(b *testing.B) {
+	var x [512]uint8 // give byte its own cache line
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.And8(&x[255], uint8(i))
+	}
+}
+
+func BenchmarkAnd(b *testing.B) {
+	var x [128]uint32 // give x its own cache line
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.And(&x[63], uint32(i))
+	}
+}
+
+func BenchmarkAnd8Parallel(b *testing.B) {
+	var x [512]uint8 // give byte its own cache line
+	sink = &x
+	b.RunParallel(func(pb *testing.PB) {
+		i := uint8(0)
+		for pb.Next() {
+			atomic.And8(&x[255], i)
+			i++
+		}
+	})
+}
+
+func BenchmarkAndParallel(b *testing.B) {
+	var x [128]uint32 // give x its own cache line
+	sink = &x
+	b.RunParallel(func(pb *testing.PB) {
+		i := uint32(0)
+		for pb.Next() {
+			atomic.And(&x[63], i)
+			i++
+		}
+	})
+}
+
+func BenchmarkOr8(b *testing.B) {
+	var x [512]uint8 // give byte its own cache line
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.Or8(&x[255], uint8(i))
+	}
+}
+
+func BenchmarkOr(b *testing.B) {
+	var x [128]uint32 // give x its own cache line
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.Or(&x[63], uint32(i))
+	}
+}
+
+func BenchmarkOr8Parallel(b *testing.B) {
+	var x [512]uint8 // give byte its own cache line
+	sink = &x
+	b.RunParallel(func(pb *testing.PB) {
+		i := uint8(0)
+		for pb.Next() {
+			atomic.Or8(&x[255], i)
+			i++
+		}
+	})
+}
+
+func BenchmarkOrParallel(b *testing.B) {
+	var x [128]uint32 // give x its own cache line
+	sink = &x
+	b.RunParallel(func(pb *testing.PB) {
+		i := uint32(0)
+		for pb.Next() {
+			atomic.Or(&x[63], i)
+			i++
+		}
+	})
+}
+
+func BenchmarkXadd(b *testing.B) {
+	var x uint32
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			atomic.Xadd(ptr, 1)
+		}
+	})
+}
+
+func BenchmarkXadd64(b *testing.B) {
+	var x uint64
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			atomic.Xadd64(ptr, 1)
+		}
+	})
+}
+
+func BenchmarkCas(b *testing.B) {
+	var x uint32
+	x = 1
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			atomic.Cas(ptr, 1, 0)
+			atomic.Cas(ptr, 0, 1)
+		}
+	})
+}
+
+func BenchmarkCas64(b *testing.B) {
+	var x uint64
+	x = 1
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		for pb.Next() {
+			atomic.Cas64(ptr, 1, 0)
+			atomic.Cas64(ptr, 0, 1)
+		}
+	})
+}
+func BenchmarkXchg(b *testing.B) {
+	var x uint32
+	x = 1
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		var y uint32
+		y = 1
+		for pb.Next() {
+			y = atomic.Xchg(ptr, y)
+			y += 1
+		}
+	})
+}
+
+func BenchmarkXchg64(b *testing.B) {
+	var x uint64
+	x = 1
+	ptr := &x
+	b.RunParallel(func(pb *testing.PB) {
+		var y uint64
+		y = 1
+		for pb.Next() {
+			y = atomic.Xchg64(ptr, y)
+			y += 1
+		}
+	})
+}
diff --git a/src/runtime/internal/atomic/doc.go b/src/runtime/internal/atomic/doc.go
new file mode 100644
index 0000000..08e6b6c
--- /dev/null
+++ b/src/runtime/internal/atomic/doc.go
@@ -0,0 +1,18 @@
+// 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.
+
+/*
+Package atomic provides atomic operations, independent of sync/atomic,
+to the runtime.
+
+On most platforms, the compiler is aware of the functions defined
+in this package, and they're replaced with platform-specific intrinsics.
+On other platforms, generic implementations are made available.
+
+Unless otherwise noted, operations defined in this package are sequentially
+consistent across threads with respect to the values they manipulate. More
+specifically, operations that happen in a specific order on one thread,
+will always be observed to happen in exactly that order by another thread.
+*/
+package atomic
diff --git a/src/runtime/internal/atomic/stubs.go b/src/runtime/internal/atomic/stubs.go
new file mode 100644
index 0000000..7df8d9c
--- /dev/null
+++ b/src/runtime/internal/atomic/stubs.go
@@ -0,0 +1,59 @@
+// Copyright 2015 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 !wasm
+
+package atomic
+
+import "unsafe"
+
+//go:noescape
+func Cas(ptr *uint32, old, new uint32) bool
+
+// NO go:noescape annotation; see atomic_pointer.go.
+func Casp1(ptr *unsafe.Pointer, old, new unsafe.Pointer) bool
+
+//go:noescape
+func Casint32(ptr *int32, old, new int32) bool
+
+//go:noescape
+func Casint64(ptr *int64, old, new int64) bool
+
+//go:noescape
+func Casuintptr(ptr *uintptr, old, new uintptr) bool
+
+//go:noescape
+func Storeint32(ptr *int32, new int32)
+
+//go:noescape
+func Storeint64(ptr *int64, new int64)
+
+//go:noescape
+func Storeuintptr(ptr *uintptr, new uintptr)
+
+//go:noescape
+func Loaduintptr(ptr *uintptr) uintptr
+
+//go:noescape
+func Loaduint(ptr *uint) uint
+
+// TODO(matloob): Should these functions have the go:noescape annotation?
+
+//go:noescape
+func Loadint32(ptr *int32) int32
+
+//go:noescape
+func Loadint64(ptr *int64) int64
+
+//go:noescape
+func Xaddint32(ptr *int32, delta int32) int32
+
+//go:noescape
+func Xaddint64(ptr *int64, delta int64) int64
+
+//go:noescape
+func Xchgint32(ptr *int32, new int32) int32
+
+//go:noescape
+func Xchgint64(ptr *int64, new int64) int64
diff --git a/src/runtime/internal/atomic/sys_linux_arm.s b/src/runtime/internal/atomic/sys_linux_arm.s
new file mode 100644
index 0000000..0cc7fa7
--- /dev/null
+++ b/src/runtime/internal/atomic/sys_linux_arm.s
@@ -0,0 +1,144 @@
+// Copyright 2015 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.
+
+#include "textflag.h"
+
+// Linux/ARM atomic operations.
+
+// Because there is so much variation in ARM devices,
+// the Linux kernel provides an appropriate compare-and-swap
+// implementation at address 0xffff0fc0.  Caller sets:
+//	R0 = old value
+//	R1 = new value
+//	R2 = addr
+//	LR = return address
+// The function returns with CS true if the swap happened.
+// http://lxr.linux.no/linux+v2.6.37.2/arch/arm/kernel/entry-armv.S#L850
+// On older kernels (before 2.6.24) the function can incorrectly
+// report a conflict, so we have to double-check the compare ourselves
+// and retry if necessary.
+//
+// https://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=b49c0f24cf6744a3f4fd09289fe7cade349dead5
+//
+TEXT cas<>(SB),NOSPLIT,$0
+	MOVW	$0xffff0fc0, R15 // R15 is hardware PC.
+
+TEXT ·Cas(SB),NOSPLIT|NOFRAME,$0
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	JMP	·armcas(SB)
+	JMP	kernelcas<>(SB)
+
+TEXT kernelcas<>(SB),NOSPLIT,$0
+	MOVW	ptr+0(FP), R2
+	// trigger potential paging fault here,
+	// because we don't know how to traceback through __kuser_cmpxchg
+	MOVW    (R2), R0
+	MOVW	old+4(FP), R0
+loop:
+	MOVW	new+8(FP), R1
+	BL	cas<>(SB)
+	BCC	check
+	MOVW	$1, R0
+	MOVB	R0, ret+12(FP)
+	RET
+check:
+	// Kernel lies; double-check.
+	MOVW	ptr+0(FP), R2
+	MOVW	old+4(FP), R0
+	MOVW	0(R2), R3
+	CMP	R0, R3
+	BEQ	loop
+	MOVW	$0, R0
+	MOVB	R0, ret+12(FP)
+	RET
+
+// As for cas, memory barriers are complicated on ARM, but the kernel
+// provides a user helper. ARMv5 does not support SMP and has no
+// memory barrier instruction at all. ARMv6 added SMP support and has
+// a memory barrier, but it requires writing to a coprocessor
+// register. ARMv7 introduced the DMB instruction, but it's expensive
+// even on single-core devices. The kernel helper takes care of all of
+// this for us.
+
+// Use kernel helper version of memory_barrier, when compiled with GOARM < 7.
+TEXT memory_barrier<>(SB),NOSPLIT|NOFRAME,$0
+	MOVW	$0xffff0fa0, R15 // R15 is hardware PC.
+
+TEXT	·Load(SB),NOSPLIT,$0-8
+	MOVW	addr+0(FP), R0
+	MOVW	(R0), R1
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BGE	native_barrier
+	BL	memory_barrier<>(SB)
+	B	end
+native_barrier:
+	DMB	MB_ISH
+end:
+	MOVW	R1, ret+4(FP)
+	RET
+
+TEXT	·Store(SB),NOSPLIT,$0-8
+	MOVW	addr+0(FP), R1
+	MOVW	v+4(FP), R2
+
+	MOVB	runtime·goarm(SB), R8
+	CMP	$7, R8
+	BGE	native_barrier
+	BL	memory_barrier<>(SB)
+	B	store
+native_barrier:
+	DMB	MB_ISH
+
+store:
+	MOVW	R2, (R1)
+
+	CMP	$7, R8
+	BGE	native_barrier2
+	BL	memory_barrier<>(SB)
+	RET
+native_barrier2:
+	DMB	MB_ISH
+	RET
+
+TEXT	·Load8(SB),NOSPLIT,$0-5
+	MOVW	addr+0(FP), R0
+	MOVB	(R0), R1
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BGE	native_barrier
+	BL	memory_barrier<>(SB)
+	B	end
+native_barrier:
+	DMB	MB_ISH
+end:
+	MOVB	R1, ret+4(FP)
+	RET
+
+TEXT	·Store8(SB),NOSPLIT,$0-5
+	MOVW	addr+0(FP), R1
+	MOVB	v+4(FP), R2
+
+	MOVB	runtime·goarm(SB), R8
+	CMP	$7, R8
+	BGE	native_barrier
+	BL	memory_barrier<>(SB)
+	B	store
+native_barrier:
+	DMB	MB_ISH
+
+store:
+	MOVB	R2, (R1)
+
+	CMP	$7, R8
+	BGE	native_barrier2
+	BL	memory_barrier<>(SB)
+	RET
+native_barrier2:
+	DMB	MB_ISH
+	RET
diff --git a/src/runtime/internal/atomic/sys_nonlinux_arm.s b/src/runtime/internal/atomic/sys_nonlinux_arm.s
new file mode 100644
index 0000000..b55bf90
--- /dev/null
+++ b/src/runtime/internal/atomic/sys_nonlinux_arm.s
@@ -0,0 +1,79 @@
+// Copyright 2015 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 !linux
+
+#include "textflag.h"
+
+// TODO(minux): this is only valid for ARMv6+
+// bool armcas(int32 *val, int32 old, int32 new)
+// Atomically:
+//	if(*val == old){
+//		*val = new;
+//		return 1;
+//	}else
+//		return 0;
+TEXT	·Cas(SB),NOSPLIT,$0
+	JMP	·armcas(SB)
+
+// Non-linux OSes support only single processor machines before ARMv7.
+// So we don't need memory barriers if goarm < 7. And we fail loud at
+// startup (runtime.checkgoarm) if it is a multi-processor but goarm < 7.
+
+TEXT	·Load(SB),NOSPLIT|NOFRAME,$0-8
+	MOVW	addr+0(FP), R0
+	MOVW	(R0), R1
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	DMB	MB_ISH
+
+	MOVW	R1, ret+4(FP)
+	RET
+
+TEXT	·Store(SB),NOSPLIT,$0-8
+	MOVW	addr+0(FP), R1
+	MOVW	v+4(FP), R2
+
+	MOVB	runtime·goarm(SB), R8
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISH
+
+	MOVW	R2, (R1)
+
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISH
+	RET
+
+TEXT	·Load8(SB),NOSPLIT|NOFRAME,$0-5
+	MOVW	addr+0(FP), R0
+	MOVB	(R0), R1
+
+	MOVB	runtime·goarm(SB), R11
+	CMP	$7, R11
+	BLT	2(PC)
+	DMB	MB_ISH
+
+	MOVB	R1, ret+4(FP)
+	RET
+
+TEXT	·Store8(SB),NOSPLIT,$0-5
+	MOVW	addr+0(FP), R1
+	MOVB	v+4(FP), R2
+
+	MOVB	runtime·goarm(SB), R8
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISH
+
+	MOVB	R2, (R1)
+
+	CMP	$7, R8
+	BLT	2(PC)
+	DMB	MB_ISH
+	RET
+
diff --git a/src/runtime/internal/atomic/types.go b/src/runtime/internal/atomic/types.go
new file mode 100644
index 0000000..d346a76
--- /dev/null
+++ b/src/runtime/internal/atomic/types.go
@@ -0,0 +1,431 @@
+// 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.
+
+package atomic
+
+import "unsafe"
+
+// Int32 is an atomically accessed int32 value.
+//
+// An Int32 must not be copied.
+type Int32 struct {
+	noCopy noCopy
+	value  int32
+}
+
+// Load accesses and returns the value atomically.
+func (i *Int32) Load() int32 {
+	return Loadint32(&i.value)
+}
+
+// Store updates the value atomically.
+func (i *Int32) Store(value int32) {
+	Storeint32(&i.value, value)
+}
+
+// CompareAndSwap atomically compares i's value with old,
+// and if they're equal, swaps i's value with new.
+//
+// Returns true if the operation succeeded.
+func (i *Int32) CompareAndSwap(old, new int32) bool {
+	return Casint32(&i.value, old, new)
+}
+
+// Swap replaces i's value with new, returning
+// i's value before the replacement.
+func (i *Int32) Swap(new int32) int32 {
+	return Xchgint32(&i.value, new)
+}
+
+// Add adds delta to i atomically, returning
+// the new updated value.
+//
+// This operation wraps around in the usual
+// two's-complement way.
+func (i *Int32) Add(delta int32) int32 {
+	return Xaddint32(&i.value, delta)
+}
+
+// Int64 is an atomically accessed int64 value.
+//
+// 8-byte aligned on all platforms, unlike a regular int64.
+//
+// An Int64 must not be copied.
+type Int64 struct {
+	noCopy noCopy
+	_      align64
+	value  int64
+}
+
+// Load accesses and returns the value atomically.
+func (i *Int64) Load() int64 {
+	return Loadint64(&i.value)
+}
+
+// Store updates the value atomically.
+func (i *Int64) Store(value int64) {
+	Storeint64(&i.value, value)
+}
+
+// CompareAndSwap atomically compares i's value with old,
+// and if they're equal, swaps i's value with new.
+//
+// Returns true if the operation succeeded.
+func (i *Int64) CompareAndSwap(old, new int64) bool {
+	return Casint64(&i.value, old, new)
+}
+
+// Swap replaces i's value with new, returning
+// i's value before the replacement.
+func (i *Int64) Swap(new int64) int64 {
+	return Xchgint64(&i.value, new)
+}
+
+// Add adds delta to i atomically, returning
+// the new updated value.
+//
+// This operation wraps around in the usual
+// two's-complement way.
+func (i *Int64) Add(delta int64) int64 {
+	return Xaddint64(&i.value, delta)
+}
+
+// Uint8 is an atomically accessed uint8 value.
+//
+// A Uint8 must not be copied.
+type Uint8 struct {
+	noCopy noCopy
+	value  uint8
+}
+
+// Load accesses and returns the value atomically.
+func (u *Uint8) Load() uint8 {
+	return Load8(&u.value)
+}
+
+// Store updates the value atomically.
+func (u *Uint8) Store(value uint8) {
+	Store8(&u.value, value)
+}
+
+// And takes value and performs a bit-wise
+// "and" operation with the value of u, storing
+// the result into u.
+//
+// The full process is performed atomically.
+func (u *Uint8) And(value uint8) {
+	And8(&u.value, value)
+}
+
+// Or takes value and performs a bit-wise
+// "or" operation with the value of u, storing
+// the result into u.
+//
+// The full process is performed atomically.
+func (u *Uint8) Or(value uint8) {
+	Or8(&u.value, value)
+}
+
+// Bool is an atomically accessed bool value.
+//
+// A Bool must not be copied.
+type Bool struct {
+	// Inherits noCopy from Uint8.
+	u Uint8
+}
+
+// Load accesses and returns the value atomically.
+func (b *Bool) Load() bool {
+	return b.u.Load() != 0
+}
+
+// Store updates the value atomically.
+func (b *Bool) Store(value bool) {
+	s := uint8(0)
+	if value {
+		s = 1
+	}
+	b.u.Store(s)
+}
+
+// Uint32 is an atomically accessed uint32 value.
+//
+// A Uint32 must not be copied.
+type Uint32 struct {
+	noCopy noCopy
+	value  uint32
+}
+
+// Load accesses and returns the value atomically.
+func (u *Uint32) Load() uint32 {
+	return Load(&u.value)
+}
+
+// LoadAcquire is a partially unsynchronized version
+// of Load that relaxes ordering constraints. Other threads
+// may observe operations that precede this operation to
+// occur after it, but no operation that occurs after it
+// on this thread can be observed to occur before it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uint32) LoadAcquire() uint32 {
+	return LoadAcq(&u.value)
+}
+
+// Store updates the value atomically.
+func (u *Uint32) Store(value uint32) {
+	Store(&u.value, value)
+}
+
+// StoreRelease is a partially unsynchronized version
+// of Store that relaxes ordering constraints. Other threads
+// may observe operations that occur after this operation to
+// precede it, but no operation that precedes it
+// on this thread can be observed to occur after it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uint32) StoreRelease(value uint32) {
+	StoreRel(&u.value, value)
+}
+
+// CompareAndSwap atomically compares u's value with old,
+// and if they're equal, swaps u's value with new.
+//
+// Returns true if the operation succeeded.
+func (u *Uint32) CompareAndSwap(old, new uint32) bool {
+	return Cas(&u.value, old, new)
+}
+
+// CompareAndSwapRelease is a partially unsynchronized version
+// of Cas that relaxes ordering constraints. Other threads
+// may observe operations that occur after this operation to
+// precede it, but no operation that precedes it
+// on this thread can be observed to occur after it.
+//
+// Returns true if the operation succeeded.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uint32) CompareAndSwapRelease(old, new uint32) bool {
+	return CasRel(&u.value, old, new)
+}
+
+// Swap replaces u's value with new, returning
+// u's value before the replacement.
+func (u *Uint32) Swap(value uint32) uint32 {
+	return Xchg(&u.value, value)
+}
+
+// And takes value and performs a bit-wise
+// "and" operation with the value of u, storing
+// the result into u.
+//
+// The full process is performed atomically.
+func (u *Uint32) And(value uint32) {
+	And(&u.value, value)
+}
+
+// Or takes value and performs a bit-wise
+// "or" operation with the value of u, storing
+// the result into u.
+//
+// The full process is performed atomically.
+func (u *Uint32) Or(value uint32) {
+	Or(&u.value, value)
+}
+
+// Add adds delta to u atomically, returning
+// the new updated value.
+//
+// This operation wraps around in the usual
+// two's-complement way.
+func (u *Uint32) Add(delta int32) uint32 {
+	return Xadd(&u.value, delta)
+}
+
+// Uint64 is an atomically accessed uint64 value.
+//
+// 8-byte aligned on all platforms, unlike a regular uint64.
+//
+// A Uint64 must not be copied.
+type Uint64 struct {
+	noCopy noCopy
+	_      align64
+	value  uint64
+}
+
+// Load accesses and returns the value atomically.
+func (u *Uint64) Load() uint64 {
+	return Load64(&u.value)
+}
+
+// Store updates the value atomically.
+func (u *Uint64) Store(value uint64) {
+	Store64(&u.value, value)
+}
+
+// CompareAndSwap atomically compares u's value with old,
+// and if they're equal, swaps u's value with new.
+//
+// Returns true if the operation succeeded.
+func (u *Uint64) CompareAndSwap(old, new uint64) bool {
+	return Cas64(&u.value, old, new)
+}
+
+// Swap replaces u's value with new, returning
+// u's value before the replacement.
+func (u *Uint64) Swap(value uint64) uint64 {
+	return Xchg64(&u.value, value)
+}
+
+// Add adds delta to u atomically, returning
+// the new updated value.
+//
+// This operation wraps around in the usual
+// two's-complement way.
+func (u *Uint64) Add(delta int64) uint64 {
+	return Xadd64(&u.value, delta)
+}
+
+// Uintptr is an atomically accessed uintptr value.
+//
+// A Uintptr must not be copied.
+type Uintptr struct {
+	noCopy noCopy
+	value  uintptr
+}
+
+// Load accesses and returns the value atomically.
+func (u *Uintptr) Load() uintptr {
+	return Loaduintptr(&u.value)
+}
+
+// LoadAcquire is a partially unsynchronized version
+// of Load that relaxes ordering constraints. Other threads
+// may observe operations that precede this operation to
+// occur after it, but no operation that occurs after it
+// on this thread can be observed to occur before it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uintptr) LoadAcquire() uintptr {
+	return LoadAcquintptr(&u.value)
+}
+
+// Store updates the value atomically.
+func (u *Uintptr) Store(value uintptr) {
+	Storeuintptr(&u.value, value)
+}
+
+// StoreRelease is a partially unsynchronized version
+// of Store that relaxes ordering constraints. Other threads
+// may observe operations that occur after this operation to
+// precede it, but no operation that precedes it
+// on this thread can be observed to occur after it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uintptr) StoreRelease(value uintptr) {
+	StoreReluintptr(&u.value, value)
+}
+
+// CompareAndSwap atomically compares u's value with old,
+// and if they're equal, swaps u's value with new.
+//
+// Returns true if the operation succeeded.
+func (u *Uintptr) CompareAndSwap(old, new uintptr) bool {
+	return Casuintptr(&u.value, old, new)
+}
+
+// Swap replaces u's value with new, returning
+// u's value before the replacement.
+func (u *Uintptr) Swap(value uintptr) uintptr {
+	return Xchguintptr(&u.value, value)
+}
+
+// Add adds delta to u atomically, returning
+// the new updated value.
+//
+// This operation wraps around in the usual
+// two's-complement way.
+func (u *Uintptr) Add(delta uintptr) uintptr {
+	return Xadduintptr(&u.value, delta)
+}
+
+// Float64 is an atomically accessed float64 value.
+//
+// 8-byte aligned on all platforms, unlike a regular float64.
+//
+// A Float64 must not be copied.
+type Float64 struct {
+	// Inherits noCopy and align64 from Uint64.
+	u Uint64
+}
+
+// Load accesses and returns the value atomically.
+func (f *Float64) Load() float64 {
+	r := f.u.Load()
+	return *(*float64)(unsafe.Pointer(&r))
+}
+
+// Store updates the value atomically.
+func (f *Float64) Store(value float64) {
+	f.u.Store(*(*uint64)(unsafe.Pointer(&value)))
+}
+
+// UnsafePointer is an atomically accessed unsafe.Pointer value.
+//
+// Note that because of the atomicity guarantees, stores to values
+// of this type never trigger a write barrier, and the relevant
+// methods are suffixed with "NoWB" to indicate that explicitly.
+// As a result, this type should be used carefully, and sparingly,
+// mostly with values that do not live in the Go heap anyway.
+//
+// An UnsafePointer must not be copied.
+type UnsafePointer struct {
+	noCopy noCopy
+	value  unsafe.Pointer
+}
+
+// Load accesses and returns the value atomically.
+func (u *UnsafePointer) Load() unsafe.Pointer {
+	return Loadp(unsafe.Pointer(&u.value))
+}
+
+// StoreNoWB updates the value atomically.
+//
+// WARNING: As the name implies this operation does *not*
+// perform a write barrier on value, and so this operation may
+// hide pointers from the GC. Use with care and sparingly.
+// It is safe to use with values not found in the Go heap.
+func (u *UnsafePointer) StoreNoWB(value unsafe.Pointer) {
+	StorepNoWB(unsafe.Pointer(&u.value), value)
+}
+
+// CompareAndSwapNoWB atomically (with respect to other methods)
+// compares u's value with old, and if they're equal,
+// swaps u's value with new.
+//
+// Returns true if the operation succeeded.
+//
+// WARNING: As the name implies this operation does *not*
+// perform a write barrier on value, and so this operation may
+// hide pointers from the GC. Use with care and sparingly.
+// It is safe to use with values not found in the Go heap.
+func (u *UnsafePointer) CompareAndSwapNoWB(old, new unsafe.Pointer) bool {
+	return Casp1(&u.value, old, new)
+}
+
+// noCopy may be embedded into structs which must not be copied
+// after the first use.
+//
+// See https://golang.org/issues/8005#issuecomment-190753527
+// for details.
+type noCopy struct{}
+
+// Lock is a no-op used by -copylocks checker from `go vet`.
+func (*noCopy) Lock()   {}
+func (*noCopy) Unlock() {}
+
+// align64 may be added to structs that must be 64-bit aligned.
+// This struct is recognized by a special case in the compiler
+// and will not work if copied to any other package.
+type align64 struct{}
diff --git a/src/runtime/internal/atomic/types_64bit.go b/src/runtime/internal/atomic/types_64bit.go
new file mode 100644
index 0000000..43c1ba2
--- /dev/null
+++ b/src/runtime/internal/atomic/types_64bit.go
@@ -0,0 +1,29 @@
+// 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.
+
+//go:build amd64 || arm64 || mips64 || mips64le || ppc64 || ppc64le || riscv64 || s390x || wasm
+
+package atomic
+
+// LoadAcquire is a partially unsynchronized version
+// of Load that relaxes ordering constraints. Other threads
+// may observe operations that precede this operation to
+// occur after it, but no operation that occurs after it
+// on this thread can be observed to occur before it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uint64) LoadAcquire() uint64 {
+	return LoadAcq64(&u.value)
+}
+
+// StoreRelease is a partially unsynchronized version
+// of Store that relaxes ordering constraints. Other threads
+// may observe operations that occur after this operation to
+// precede it, but no operation that precedes it
+// on this thread can be observed to occur after it.
+//
+// WARNING: Use sparingly and with great care.
+func (u *Uint64) StoreRelease(value uint64) {
+	StoreRel64(&u.value, value)
+}
diff --git a/src/runtime/internal/atomic/unaligned.go b/src/runtime/internal/atomic/unaligned.go
new file mode 100644
index 0000000..a859de4
--- /dev/null
+++ b/src/runtime/internal/atomic/unaligned.go
@@ -0,0 +1,9 @@
+// Copyright 2020 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 atomic
+
+func panicUnaligned() {
+	panic("unaligned 64-bit atomic operation")
+}
diff --git a/src/runtime/internal/math/math.go b/src/runtime/internal/math/math.go
new file mode 100644
index 0000000..c3fac36
--- /dev/null
+++ b/src/runtime/internal/math/math.go
@@ -0,0 +1,40 @@
+// 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 math
+
+import "internal/goarch"
+
+const MaxUintptr = ^uintptr(0)
+
+// MulUintptr returns a * b and whether the multiplication overflowed.
+// On supported platforms this is an intrinsic lowered by the compiler.
+func MulUintptr(a, b uintptr) (uintptr, bool) {
+	if a|b < 1<<(4*goarch.PtrSize) || a == 0 {
+		return a * b, false
+	}
+	overflow := b > MaxUintptr/a
+	return a * b, overflow
+}
+
+// Mul64 returns the 128-bit product of x and y: (hi, lo) = x * y
+// with the product bits' upper half returned in hi and the lower
+// half returned in lo.
+// This is a copy from math/bits.Mul64
+// On supported platforms this is an intrinsic lowered by the compiler.
+func Mul64(x, y uint64) (hi, lo uint64) {
+	const mask32 = 1<<32 - 1
+	x0 := x & mask32
+	x1 := x >> 32
+	y0 := y & mask32
+	y1 := y >> 32
+	w0 := x0 * y0
+	t := x1*y0 + w0>>32
+	w1 := t & mask32
+	w2 := t >> 32
+	w1 += x0 * y1
+	hi = x1*y1 + w2 + w1>>32
+	lo = x * y
+	return
+}
diff --git a/src/runtime/internal/math/math_test.go b/src/runtime/internal/math/math_test.go
new file mode 100644
index 0000000..303eb63
--- /dev/null
+++ b/src/runtime/internal/math/math_test.go
@@ -0,0 +1,79 @@
+// 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 math_test
+
+import (
+	. "runtime/internal/math"
+	"testing"
+)
+
+const (
+	UintptrSize = 32 << (^uintptr(0) >> 63)
+)
+
+type mulUintptrTest struct {
+	a        uintptr
+	b        uintptr
+	overflow bool
+}
+
+var mulUintptrTests = []mulUintptrTest{
+	{0, 0, false},
+	{1000, 1000, false},
+	{MaxUintptr, 0, false},
+	{MaxUintptr, 1, false},
+	{MaxUintptr / 2, 2, false},
+	{MaxUintptr / 2, 3, true},
+	{MaxUintptr, 10, true},
+	{MaxUintptr, 100, true},
+	{MaxUintptr / 100, 100, false},
+	{MaxUintptr / 1000, 1001, true},
+	{1<<(UintptrSize/2) - 1, 1<<(UintptrSize/2) - 1, false},
+	{1 << (UintptrSize / 2), 1 << (UintptrSize / 2), true},
+	{MaxUintptr >> 32, MaxUintptr >> 32, false},
+	{MaxUintptr, MaxUintptr, true},
+}
+
+func TestMulUintptr(t *testing.T) {
+	for _, test := range mulUintptrTests {
+		a, b := test.a, test.b
+		for i := 0; i < 2; i++ {
+			mul, overflow := MulUintptr(a, b)
+			if mul != a*b || overflow != test.overflow {
+				t.Errorf("MulUintptr(%v, %v) = %v, %v want %v, %v",
+					a, b, mul, overflow, a*b, test.overflow)
+			}
+			a, b = b, a
+		}
+	}
+}
+
+var SinkUintptr uintptr
+var SinkBool bool
+
+var x, y uintptr
+
+func BenchmarkMulUintptr(b *testing.B) {
+	x, y = 1, 2
+	b.Run("small", func(b *testing.B) {
+		for i := 0; i < b.N; i++ {
+			var overflow bool
+			SinkUintptr, overflow = MulUintptr(x, y)
+			if overflow {
+				SinkUintptr = 0
+			}
+		}
+	})
+	x, y = MaxUintptr, MaxUintptr-1
+	b.Run("large", func(b *testing.B) {
+		for i := 0; i < b.N; i++ {
+			var overflow bool
+			SinkUintptr, overflow = MulUintptr(x, y)
+			if overflow {
+				SinkUintptr = 0
+			}
+		}
+	})
+}
diff --git a/src/runtime/internal/sys/consts.go b/src/runtime/internal/sys/consts.go
new file mode 100644
index 0000000..fffcf81
--- /dev/null
+++ b/src/runtime/internal/sys/consts.go
@@ -0,0 +1,34 @@
+// Copyright 2014 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 sys
+
+import (
+	"internal/goarch"
+	"internal/goos"
+)
+
+// AIX requires a larger stack for syscalls.
+const StackGuardMultiplier = StackGuardMultiplierDefault*(1-goos.IsAix) + 2*goos.IsAix
+
+// DefaultPhysPageSize is the default physical page size.
+const DefaultPhysPageSize = goarch.DefaultPhysPageSize
+
+// PCQuantum is the minimal unit for a program counter (1 on x86, 4 on most other systems).
+// The various PC tables record PC deltas pre-divided by PCQuantum.
+const PCQuantum = goarch.PCQuantum
+
+// Int64Align is the required alignment for a 64-bit integer (4 on 32-bit systems, 8 on 64-bit).
+const Int64Align = goarch.PtrSize
+
+// MinFrameSize is the size of the system-reserved words at the bottom
+// of a frame (just above the architectural stack pointer).
+// It is zero on x86 and PtrSize on most non-x86 (LR-based) systems.
+// On PowerPC it is larger, to cover three more reserved words:
+// the compiler word, the link editor word, and the TOC save word.
+const MinFrameSize = goarch.MinFrameSize
+
+// StackAlign is the required alignment of the SP register.
+// The stack must be at least word aligned, but some architectures require more.
+const StackAlign = goarch.StackAlign
diff --git a/src/runtime/internal/sys/intrinsics.go b/src/runtime/internal/sys/intrinsics.go
new file mode 100644
index 0000000..5af4901
--- /dev/null
+++ b/src/runtime/internal/sys/intrinsics.go
@@ -0,0 +1,91 @@
+// Copyright 2016 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
+
+// TODO finish intrinsifying 386, deadcode the assembly, remove build tags, merge w/ intrinsics_common
+// TODO replace all uses of CtzXX with TrailingZerosXX; they are the same.
+
+package sys
+
+// Using techniques from http://supertech.csail.mit.edu/papers/debruijn.pdf
+
+const deBruijn64ctz = 0x0218a392cd3d5dbf
+
+var deBruijnIdx64ctz = [64]byte{
+	0, 1, 2, 7, 3, 13, 8, 19,
+	4, 25, 14, 28, 9, 34, 20, 40,
+	5, 17, 26, 38, 15, 46, 29, 48,
+	10, 31, 35, 54, 21, 50, 41, 57,
+	63, 6, 12, 18, 24, 27, 33, 39,
+	16, 37, 45, 47, 30, 53, 49, 56,
+	62, 11, 23, 32, 36, 44, 52, 55,
+	61, 22, 43, 51, 60, 42, 59, 58,
+}
+
+const deBruijn32ctz = 0x04653adf
+
+var deBruijnIdx32ctz = [32]byte{
+	0, 1, 2, 6, 3, 11, 7, 16,
+	4, 14, 12, 21, 8, 23, 17, 26,
+	31, 5, 10, 15, 13, 20, 22, 25,
+	30, 9, 19, 24, 29, 18, 28, 27,
+}
+
+// Ctz64 counts trailing (low-order) zeroes,
+// and if all are zero, then 64.
+func Ctz64(x uint64) int {
+	x &= -x                       // isolate low-order bit
+	y := x * deBruijn64ctz >> 58  // extract part of deBruijn sequence
+	i := int(deBruijnIdx64ctz[y]) // convert to bit index
+	z := int((x - 1) >> 57 & 64)  // adjustment if zero
+	return i + z
+}
+
+// Ctz32 counts trailing (low-order) zeroes,
+// and if all are zero, then 32.
+func Ctz32(x uint32) int {
+	x &= -x                       // isolate low-order bit
+	y := x * deBruijn32ctz >> 27  // extract part of deBruijn sequence
+	i := int(deBruijnIdx32ctz[y]) // convert to bit index
+	z := int((x - 1) >> 26 & 32)  // adjustment if zero
+	return i + z
+}
+
+// Ctz8 returns the number of trailing zero bits in x; the result is 8 for x == 0.
+func Ctz8(x uint8) int {
+	return int(ntz8tab[x])
+}
+
+// Bswap64 returns its input with byte order reversed
+// 0x0102030405060708 -> 0x0807060504030201
+func Bswap64(x uint64) uint64 {
+	c8 := uint64(0x00ff00ff00ff00ff)
+	a := x >> 8 & c8
+	b := (x & c8) << 8
+	x = a | b
+	c16 := uint64(0x0000ffff0000ffff)
+	a = x >> 16 & c16
+	b = (x & c16) << 16
+	x = a | b
+	c32 := uint64(0x00000000ffffffff)
+	a = x >> 32 & c32
+	b = (x & c32) << 32
+	x = a | b
+	return x
+}
+
+// Bswap32 returns its input with byte order reversed
+// 0x01020304 -> 0x04030201
+func Bswap32(x uint32) uint32 {
+	c8 := uint32(0x00ff00ff)
+	a := x >> 8 & c8
+	b := (x & c8) << 8
+	x = a | b
+	c16 := uint32(0x0000ffff)
+	a = x >> 16 & c16
+	b = (x & c16) << 16
+	x = a | b
+	return x
+}
diff --git a/src/runtime/internal/sys/intrinsics_386.s b/src/runtime/internal/sys/intrinsics_386.s
new file mode 100644
index 0000000..784b246
--- /dev/null
+++ b/src/runtime/internal/sys/intrinsics_386.s
@@ -0,0 +1,58 @@
+// Copyright 2016 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.
+
+#include "textflag.h"
+
+TEXT runtime∕internal∕sys·Ctz64(SB), NOSPLIT, $0-12
+	// Try low 32 bits.
+	MOVL	x_lo+0(FP), AX
+	BSFL	AX, AX
+	JZ	tryhigh
+	MOVL	AX, ret+8(FP)
+	RET
+
+tryhigh:
+	// Try high 32 bits.
+	MOVL	x_hi+4(FP), AX
+	BSFL	AX, AX
+	JZ	none
+	ADDL	$32, AX
+	MOVL	AX, ret+8(FP)
+	RET
+
+none:
+	// No bits are set.
+	MOVL	$64, ret+8(FP)
+	RET
+
+TEXT runtime∕internal∕sys·Ctz32(SB), NOSPLIT, $0-8
+	MOVL	x+0(FP), AX
+	BSFL	AX, AX
+	JNZ	2(PC)
+	MOVL	$32, AX
+	MOVL	AX, ret+4(FP)
+	RET
+
+TEXT runtime∕internal∕sys·Ctz8(SB), NOSPLIT, $0-8
+	MOVBLZX	x+0(FP), AX
+	BSFL	AX, AX
+	JNZ	2(PC)
+	MOVL	$8, AX
+	MOVL	AX, ret+4(FP)
+	RET
+
+TEXT runtime∕internal∕sys·Bswap64(SB), NOSPLIT, $0-16
+	MOVL	x_lo+0(FP), AX
+	MOVL	x_hi+4(FP), BX
+	BSWAPL	AX
+	BSWAPL	BX
+	MOVL	BX, ret_lo+8(FP)
+	MOVL	AX, ret_hi+12(FP)
+	RET
+
+TEXT runtime∕internal∕sys·Bswap32(SB), NOSPLIT, $0-8
+	MOVL	x+0(FP), AX
+	BSWAPL	AX
+	MOVL	AX, ret+4(FP)
+	RET
diff --git a/src/runtime/internal/sys/intrinsics_common.go b/src/runtime/internal/sys/intrinsics_common.go
new file mode 100644
index 0000000..48d9759
--- /dev/null
+++ b/src/runtime/internal/sys/intrinsics_common.go
@@ -0,0 +1,158 @@
+// Copyright 2019 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 sys
+
+// Copied from math/bits to avoid dependence.
+
+var len8tab = [256]uint8{
+	0x00, 0x01, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+	0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05,
+	0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+	0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+	0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+	0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+	0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+	0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+	0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+}
+
+var ntz8tab = [256]uint8{
+	0x08, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x05, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x06, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x05, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x07, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x05, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x06, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x05, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+	0x04, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00, 0x03, 0x00, 0x01, 0x00, 0x02, 0x00, 0x01, 0x00,
+}
+
+// len64 returns the minimum number of bits required to represent x; the result is 0 for x == 0.
+func Len64(x uint64) (n int) {
+	if x >= 1<<32 {
+		x >>= 32
+		n = 32
+	}
+	if x >= 1<<16 {
+		x >>= 16
+		n += 16
+	}
+	if x >= 1<<8 {
+		x >>= 8
+		n += 8
+	}
+	return n + int(len8tab[x])
+}
+
+// --- OnesCount ---
+
+const m0 = 0x5555555555555555 // 01010101 ...
+const m1 = 0x3333333333333333 // 00110011 ...
+const m2 = 0x0f0f0f0f0f0f0f0f // 00001111 ...
+
+// OnesCount64 returns the number of one bits ("population count") in x.
+func OnesCount64(x uint64) int {
+	// Implementation: Parallel summing of adjacent bits.
+	// See "Hacker's Delight", Chap. 5: Counting Bits.
+	// The following pattern shows the general approach:
+	//
+	//   x = x>>1&(m0&m) + x&(m0&m)
+	//   x = x>>2&(m1&m) + x&(m1&m)
+	//   x = x>>4&(m2&m) + x&(m2&m)
+	//   x = x>>8&(m3&m) + x&(m3&m)
+	//   x = x>>16&(m4&m) + x&(m4&m)
+	//   x = x>>32&(m5&m) + x&(m5&m)
+	//   return int(x)
+	//
+	// Masking (& operations) can be left away when there's no
+	// danger that a field's sum will carry over into the next
+	// field: Since the result cannot be > 64, 8 bits is enough
+	// and we can ignore the masks for the shifts by 8 and up.
+	// Per "Hacker's Delight", the first line can be simplified
+	// more, but it saves at best one instruction, so we leave
+	// it alone for clarity.
+	const m = 1<<64 - 1
+	x = x>>1&(m0&m) + x&(m0&m)
+	x = x>>2&(m1&m) + x&(m1&m)
+	x = (x>>4 + x) & (m2 & m)
+	x += x >> 8
+	x += x >> 16
+	x += x >> 32
+	return int(x) & (1<<7 - 1)
+}
+
+var deBruijn64tab = [64]byte{
+	0, 1, 56, 2, 57, 49, 28, 3, 61, 58, 42, 50, 38, 29, 17, 4,
+	62, 47, 59, 36, 45, 43, 51, 22, 53, 39, 33, 30, 24, 18, 12, 5,
+	63, 55, 48, 27, 60, 41, 37, 16, 46, 35, 44, 21, 52, 32, 23, 11,
+	54, 26, 40, 15, 34, 20, 31, 10, 25, 14, 19, 9, 13, 8, 7, 6,
+}
+
+const deBruijn64 = 0x03f79d71b4ca8b09
+
+// TrailingZeros64 returns the number of trailing zero bits in x; the result is 64 for x == 0.
+func TrailingZeros64(x uint64) int {
+	if x == 0 {
+		return 64
+	}
+	// If popcount is fast, replace code below with return popcount(^x & (x - 1)).
+	//
+	// x & -x leaves only the right-most bit set in the word. Let k be the
+	// index of that bit. Since only a single bit is set, the value is two
+	// to the power of k. Multiplying by a power of two is equivalent to
+	// left shifting, in this case by k bits. The de Bruijn (64 bit) constant
+	// is such that all six bit, consecutive substrings are distinct.
+	// Therefore, if we have a left shifted version of this constant we can
+	// find by how many bits it was shifted by looking at which six bit
+	// substring ended up at the top of the word.
+	// (Knuth, volume 4, section 7.3.1)
+	return int(deBruijn64tab[(x&-x)*deBruijn64>>(64-6)])
+}
+
+// LeadingZeros64 returns the number of leading zero bits in x; the result is 64 for x == 0.
+func LeadingZeros64(x uint64) int { return 64 - Len64(x) }
+
+// LeadingZeros8 returns the number of leading zero bits in x; the result is 8 for x == 0.
+func LeadingZeros8(x uint8) int { return 8 - Len8(x) }
+
+// TrailingZeros8 returns the number of trailing zero bits in x; the result is 8 for x == 0.
+func TrailingZeros8(x uint8) int {
+	return int(ntz8tab[x])
+}
+
+// Len8 returns the minimum number of bits required to represent x; the result is 0 for x == 0.
+func Len8(x uint8) int {
+	return int(len8tab[x])
+}
+
+// Prefetch prefetches data from memory addr to cache
+//
+// AMD64: Produce PREFETCHT0 instruction
+//
+// ARM64: Produce PRFM instruction with PLDL1KEEP option
+func Prefetch(addr uintptr) {}
+
+// PrefetchStreamed prefetches data from memory addr, with a hint that this data is being streamed.
+// That is, it is likely to be accessed very soon, but only once. If possible, this will avoid polluting the cache.
+//
+// AMD64: Produce PREFETCHNTA instruction
+//
+// ARM64: Produce PRFM instruction with PLDL1STRM option
+func PrefetchStreamed(addr uintptr) {}
diff --git a/src/runtime/internal/sys/intrinsics_stubs.go b/src/runtime/internal/sys/intrinsics_stubs.go
new file mode 100644
index 0000000..a020652
--- /dev/null
+++ b/src/runtime/internal/sys/intrinsics_stubs.go
@@ -0,0 +1,13 @@
+// Copyright 2016 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
+
+package sys
+
+func Ctz64(x uint64) int
+func Ctz32(x uint32) int
+func Ctz8(x uint8) int
+func Bswap64(x uint64) uint64
+func Bswap32(x uint32) uint32
diff --git a/src/runtime/internal/sys/intrinsics_test.go b/src/runtime/internal/sys/intrinsics_test.go
new file mode 100644
index 0000000..0444183
--- /dev/null
+++ b/src/runtime/internal/sys/intrinsics_test.go
@@ -0,0 +1,38 @@
+package sys_test
+
+import (
+	"runtime/internal/sys"
+	"testing"
+)
+
+func TestCtz64(t *testing.T) {
+	for i := 0; i <= 64; i++ {
+		x := uint64(5) << uint(i)
+		if got := sys.Ctz64(x); got != i {
+			t.Errorf("Ctz64(%d)=%d, want %d", x, got, i)
+		}
+	}
+}
+func TestCtz32(t *testing.T) {
+	for i := 0; i <= 32; i++ {
+		x := uint32(5) << uint(i)
+		if got := sys.Ctz32(x); got != i {
+			t.Errorf("Ctz32(%d)=%d, want %d", x, got, i)
+		}
+	}
+}
+
+func TestBswap64(t *testing.T) {
+	x := uint64(0x1122334455667788)
+	y := sys.Bswap64(x)
+	if y != 0x8877665544332211 {
+		t.Errorf("Bswap(%x)=%x, want 0x8877665544332211", x, y)
+	}
+}
+func TestBswap32(t *testing.T) {
+	x := uint32(0x11223344)
+	y := sys.Bswap32(x)
+	if y != 0x44332211 {
+		t.Errorf("Bswap(%x)=%x, want 0x44332211", x, y)
+	}
+}
diff --git a/src/runtime/internal/sys/sys.go b/src/runtime/internal/sys/sys.go
new file mode 100644
index 0000000..694101d
--- /dev/null
+++ b/src/runtime/internal/sys/sys.go
@@ -0,0 +1,7 @@
+// Copyright 2015 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 sys contains system- and configuration- and architecture-specific
+// constants used by the runtime.
+package sys
diff --git a/src/runtime/internal/syscall/asm_linux_386.s b/src/runtime/internal/syscall/asm_linux_386.s
new file mode 100644
index 0000000..15aae4d
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_386.s
@@ -0,0 +1,34 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// See ../sys_linux_386.s for the reason why we always use int 0x80
+// instead of the glibc-specific "CALL 0x10(GS)".
+#define INVOKE_SYSCALL	INT	$0x80
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+//
+// Syscall # in AX, args in BX CX DX SI DI BP, return in AX
+TEXT ·Syscall6(SB),NOSPLIT,$0-40
+	MOVL	num+0(FP), AX	// syscall entry
+	MOVL	a1+4(FP), BX
+	MOVL	a2+8(FP), CX
+	MOVL	a3+12(FP), DX
+	MOVL	a4+16(FP), SI
+	MOVL	a5+20(FP), DI
+	MOVL	a6+24(FP), BP
+	INVOKE_SYSCALL
+	CMPL	AX, $0xfffff001
+	JLS	ok
+	MOVL	$-1, r1+28(FP)
+	MOVL	$0, r2+32(FP)
+	NEGL	AX
+	MOVL	AX, errno+36(FP)
+	RET
+ok:
+	MOVL	AX, r1+28(FP)
+	MOVL	DX, r2+32(FP)
+	MOVL	$0, errno+36(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_amd64.s b/src/runtime/internal/syscall/asm_linux_amd64.s
new file mode 100644
index 0000000..3740ef1
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_amd64.s
@@ -0,0 +1,47 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+//
+// We need to convert to the syscall ABI.
+//
+// arg | ABIInternal | Syscall
+// ---------------------------
+// num | AX          | AX
+// a1  | BX          | DI
+// a2  | CX          | SI
+// a3  | DI          | DX
+// a4  | SI          | R10
+// a5  | R8          | R8
+// a6  | R9          | R9
+//
+// r1  | AX          | AX
+// r2  | BX          | DX
+// err | CX          | part of AX
+//
+// Note that this differs from "standard" ABI convention, which would pass 4th
+// arg in CX, not R10.
+TEXT ·Syscall6<ABIInternal>(SB),NOSPLIT,$0
+	// a6 already in R9.
+	// a5 already in R8.
+	MOVQ	SI, R10 // a4
+	MOVQ	DI, DX  // a3
+	MOVQ	CX, SI  // a2
+	MOVQ	BX, DI  // a1
+	// num already in AX.
+	SYSCALL
+	CMPQ	AX, $0xfffffffffffff001
+	JLS	ok
+	NEGQ	AX
+	MOVQ	AX, CX  // errno
+	MOVQ	$-1, AX // r1
+	MOVQ	$0, BX  // r2
+	RET
+ok:
+	// r1 already in AX.
+	MOVQ	DX, BX // r2
+	MOVQ	$0, CX // errno
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_arm.s b/src/runtime/internal/syscall/asm_linux_arm.s
new file mode 100644
index 0000000..dbf1826
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_arm.s
@@ -0,0 +1,32 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-40
+	MOVW	num+0(FP), R7	// syscall entry
+	MOVW	a1+4(FP), R0
+	MOVW	a2+8(FP), R1
+	MOVW	a3+12(FP), R2
+	MOVW	a4+16(FP), R3
+	MOVW	a5+20(FP), R4
+	MOVW	a6+24(FP), R5
+	SWI	$0
+	MOVW	$0xfffff001, R6
+	CMP	R6, R0
+	BLS	ok
+	MOVW	$-1, R1
+	MOVW	R1, r1+28(FP)
+	MOVW	$0, R2
+	MOVW	R2, r2+32(FP)
+	RSB	$0, R0, R0
+	MOVW	R0, errno+36(FP)
+	RET
+ok:
+	MOVW	R0, r1+28(FP)
+	MOVW	R1, r2+32(FP)
+	MOVW	$0, R0
+	MOVW	R0, errno+36(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_arm64.s b/src/runtime/internal/syscall/asm_linux_arm64.s
new file mode 100644
index 0000000..83e862f
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_arm64.s
@@ -0,0 +1,29 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOVD	num+0(FP), R8	// syscall entry
+	MOVD	a1+8(FP), R0
+	MOVD	a2+16(FP), R1
+	MOVD	a3+24(FP), R2
+	MOVD	a4+32(FP), R3
+	MOVD	a5+40(FP), R4
+	MOVD	a6+48(FP), R5
+	SVC
+	CMN	$4095, R0
+	BCC	ok
+	MOVD	$-1, R4
+	MOVD	R4, r1+56(FP)
+	MOVD	ZR, r2+64(FP)
+	NEG	R0, R0
+	MOVD	R0, errno+72(FP)
+	RET
+ok:
+	MOVD	R0, r1+56(FP)
+	MOVD	R1, r2+64(FP)
+	MOVD	ZR, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_loong64.s b/src/runtime/internal/syscall/asm_linux_loong64.s
new file mode 100644
index 0000000..d6a33f9
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_loong64.s
@@ -0,0 +1,29 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOVV	num+0(FP), R11  // syscall entry
+	MOVV	a1+8(FP), R4
+	MOVV	a2+16(FP), R5
+	MOVV	a3+24(FP), R6
+	MOVV	a4+32(FP), R7
+	MOVV	a5+40(FP), R8
+	MOVV	a6+48(FP), R9
+	SYSCALL
+	MOVW	$-4096, R12
+	BGEU	R12, R4, ok
+	MOVV	$-1, R12
+	MOVV	R12, r1+56(FP)
+	MOVV	R0, r2+64(FP)
+	SUBVU	R4, R0, R4
+	MOVV	R4, errno+72(FP)
+	RET
+ok:
+	MOVV	R4, r1+56(FP)
+	MOVV	R0, r2+64(FP)	// r2 is not used. Always set to 0.
+	MOVV	R0, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_mips64x.s b/src/runtime/internal/syscall/asm_linux_mips64x.s
new file mode 100644
index 0000000..0e88a2d
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_mips64x.s
@@ -0,0 +1,29 @@
+// Copyright 2022 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 linux && (mips64 || mips64le)
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOVV	num+0(FP), R2	// syscall entry
+	MOVV	a1+8(FP), R4
+	MOVV	a2+16(FP), R5
+	MOVV	a3+24(FP), R6
+	MOVV	a4+32(FP), R7
+	MOVV	a5+40(FP), R8
+	MOVV	a6+48(FP), R9
+	SYSCALL
+	BEQ	R7, ok
+	MOVV	$-1, R1
+	MOVV	R1, r1+56(FP)
+	MOVV	R0, r2+64(FP)
+	MOVV	R2, errno+72(FP)
+	RET
+ok:
+	MOVV	R2, r1+56(FP)
+	MOVV	R3, r2+64(FP)
+	MOVV	R0, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_mipsx.s b/src/runtime/internal/syscall/asm_linux_mipsx.s
new file mode 100644
index 0000000..050029e
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_mipsx.s
@@ -0,0 +1,34 @@
+// Copyright 2022 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 linux && (mips || mipsle)
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+//
+// The 5th and 6th arg go at sp+16, sp+20.
+// Note that frame size of 20 means that 24 bytes gets reserved on stack.
+TEXT ·Syscall6(SB),NOSPLIT,$20-40
+	MOVW	num+0(FP), R2	// syscall entry
+	MOVW	a1+4(FP), R4
+	MOVW	a2+8(FP), R5
+	MOVW	a3+12(FP), R6
+	MOVW	a4+16(FP), R7
+	MOVW	a5+20(FP), R8
+	MOVW	a6+24(FP), R9
+	MOVW	R8, 16(R29)
+	MOVW	R9, 20(R29)
+	SYSCALL
+	BEQ	R7, ok
+	MOVW	$-1, R1
+	MOVW	R1, r1+28(FP)
+	MOVW	R0, r2+32(FP)
+	MOVW	R2, errno+36(FP)
+	RET
+ok:
+	MOVW	R2, r1+28(FP)
+	MOVW	R3, r2+32(FP)
+	MOVW	R0, errno+36(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_ppc64x.s b/src/runtime/internal/syscall/asm_linux_ppc64x.s
new file mode 100644
index 0000000..8cf8737
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_ppc64x.s
@@ -0,0 +1,28 @@
+// Copyright 2022 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 linux && (ppc64 || ppc64le)
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOVD	num+0(FP), R9	// syscall entry
+	MOVD	a1+8(FP), R3
+	MOVD	a2+16(FP), R4
+	MOVD	a3+24(FP), R5
+	MOVD	a4+32(FP), R6
+	MOVD	a5+40(FP), R7
+	MOVD	a6+48(FP), R8
+	SYSCALL	R9
+	MOVD	R0, r2+64(FP) // r2 is not used. Always set to 0.
+	BVC	ok
+	MOVD	$-1, R4
+	MOVD	R4, r1+56(FP)
+	MOVD	R3, errno+72(FP)
+	RET
+ok:
+	MOVD	R3, r1+56(FP)
+	MOVD	R0, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_riscv64.s b/src/runtime/internal/syscall/asm_linux_riscv64.s
new file mode 100644
index 0000000..a8652fd
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_riscv64.s
@@ -0,0 +1,29 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOV	num+0(FP), A7	// syscall entry
+	MOV	a1+8(FP), A0
+	MOV	a2+16(FP), A1
+	MOV	a3+24(FP), A2
+	MOV	a4+32(FP), A3
+	MOV	a5+40(FP), A4
+	MOV	a6+48(FP), A5
+	ECALL
+	MOV	$-4096, T0
+	BLTU	T0, A0, err
+	MOV	A0, r1+56(FP)
+	MOV	A1, r2+64(FP)
+	MOV	ZERO, errno+72(FP)
+	RET
+err:
+	MOV	$-1, T0
+	MOV	T0, r1+56(FP)
+	MOV	ZERO, r2+64(FP)
+	SUB	A0, ZERO, A0
+	MOV	A0, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/asm_linux_s390x.s b/src/runtime/internal/syscall/asm_linux_s390x.s
new file mode 100644
index 0000000..1b27f29
--- /dev/null
+++ b/src/runtime/internal/syscall/asm_linux_s390x.s
@@ -0,0 +1,28 @@
+// Copyright 2022 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.
+
+#include "textflag.h"
+
+// func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+TEXT ·Syscall6(SB),NOSPLIT,$0-80
+	MOVD	num+0(FP), R1	// syscall entry
+	MOVD	a1+8(FP), R2
+	MOVD	a2+16(FP), R3
+	MOVD	a3+24(FP), R4
+	MOVD	a4+32(FP), R5
+	MOVD	a5+40(FP), R6
+	MOVD	a6+48(FP), R7
+	SYSCALL
+	MOVD	$0xfffffffffffff001, R8
+	CMPUBLT	R2, R8, ok
+	MOVD	$-1, r1+56(FP)
+	MOVD	$0, r2+64(FP)
+	NEG	R2, R2
+	MOVD	R2, errno+72(FP)
+	RET
+ok:
+	MOVD	R2, r1+56(FP)
+	MOVD	R3, r2+64(FP)
+	MOVD	$0, errno+72(FP)
+	RET
diff --git a/src/runtime/internal/syscall/syscall_linux.go b/src/runtime/internal/syscall/syscall_linux.go
new file mode 100644
index 0000000..7f268e8
--- /dev/null
+++ b/src/runtime/internal/syscall/syscall_linux.go
@@ -0,0 +1,39 @@
+// Copyright 2022 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 syscall provides the syscall primitives required for the runtime.
+package syscall
+
+import (
+	_ "unsafe" // for go:linkname
+)
+
+// TODO(https://go.dev/issue/51087): This package is incomplete and currently
+// only contains very minimal support for Linux.
+
+// Syscall6 calls system call number 'num' with arguments a1-6.
+func Syscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr)
+
+// syscall_RawSyscall6 is a push linkname to export Syscall6 as
+// syscall.RawSyscall6.
+//
+// //go:uintptrkeepalive because the uintptr argument may be converted pointers
+// that need to be kept alive in the caller (this is implied for Syscall6 since
+// it has no body).
+//
+// //go:nosplit because stack copying does not account for uintptrkeepalive, so
+// the stack must not grow. Stack copying cannot blindly assume that all
+// uintptr arguments are pointers, because some values may look like pointers,
+// but not really be pointers, and adjusting their value would break the call.
+//
+// This is a separate wrapper because we can't export one function as two
+// names. The assembly implementations name themselves Syscall6 would not be
+// affected by a linkname.
+//
+//go:uintptrkeepalive
+//go:nosplit
+//go:linkname syscall_RawSyscall6 syscall.RawSyscall6
+func syscall_RawSyscall6(num, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, errno uintptr) {
+	return Syscall6(num, a1, a2, a3, a4, a5, a6)
+}