internal/runtime/atomic: add Xchg8 for arm

For #69735

Change-Id: I18c0ca15d94a9b1751c1e55459283e01dc114150
GitHub-Last-Rev: dd9a39a555
GitHub-Pull-Request: golang/go#69924
Cq-Include-Trybots: luci.golang.try:gotip-linux-arm
Reviewed-on: https://go-review.googlesource.com/c/go/+/620855
Reviewed-by: Keith Randall <khr@golang.org>
Reviewed-by: Keith Randall <khr@google.com>
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Michael Pratt <mpratt@google.com>

src/internal/runtime/atomic/atomic_arm.go

@@ -74,6 +74,27 @@ func Xchg(addr *uint32, v uint32) uint32 {
 	}
 }
 
+//go:noescape
+func Xchg8(addr *uint8, v uint8) uint8
+
+//go:nosplit
+func goXchg8(addr *uint8, v uint8) uint8 {
+	// Align down to 4 bytes and use 32-bit CAS.
+	addr32 := (*uint32)(unsafe.Pointer(uintptr(unsafe.Pointer(addr)) &^ 3))
+	shift := (uintptr(unsafe.Pointer(addr)) & 3) * 8 // little endian
+	word := uint32(v) << shift
+	mask := uint32(0xFF) << shift
+
+	for {
+		old := *addr32 // Read the old 32-bit value
+		// Clear the old 8 bits then insert the new value
+		if Cas(addr32, old, (old&^mask)|word) {
+			// Return the old 8-bit value
+			return uint8((old & mask) >> shift)
+		}
+	}
+}
+
 //go:nosplit
 func Xchguintptr(addr *uintptr, v uintptr) uintptr {
 	return uintptr(Xchg((*uint32)(unsafe.Pointer(addr)), uint32(v)))

src/internal/runtime/atomic/atomic_arm.s

@@ -264,6 +264,23 @@ or8loop:
 
 	RET
 
+TEXT armXchg8<>(SB),NOSPLIT,$0-9
+	// addr is already in R1
+	MOVB	v+4(FP), R2
+xchg8loop:
+	LDREXB	(R1), R6
+
+	DMB	MB_ISHST
+
+	STREXB	R2, (R1), R0
+	CMP	$0, R0
+	BNE	xchg8loop
+
+	DMB	MB_ISH
+
+	MOVB R6, ret+8(FP)
+	RET
+
 // The following functions all panic if their address argument isn't
 // 8-byte aligned. Since we're calling back into Go code to do this,
 // we have to cooperate with stack unwinding. In the normal case, the
@@ -374,3 +391,17 @@ TEXT ·Or8(SB),NOSPLIT,$-4-5
 	JMP	·goOr8(SB)
 #endif
 	JMP	armOr8<>(SB)
+
+TEXT ·Xchg8(SB),NOSPLIT,$-4-9
+	NO_LOCAL_POINTERS
+	MOVW	addr+0(FP), R1
+
+	// Uses STREXB/LDREXB that is armv6k or later.
+	// For simplicity we only enable this on armv7.
+#ifndef GOARM_7
+	MOVB	internal∕cpu·ARM+const_offsetARMHasV7Atomics(SB), R11
+	CMP	$1, R11
+	BEQ	2(PC)
+	JMP	·goXchg8(SB)
+#endif
+	JMP	armXchg8<>(SB)

src/internal/runtime/atomic/xchg8_test.go

@@ -2,7 +2,7 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-//go:build 386 || amd64 || arm64 || ppc64 || ppc64le
+//go:build 386 || amd64 || arm || arm64 || ppc64 || ppc64le
 
 package atomic_test
 
@@ -37,3 +37,23 @@ func TestXchg8(t *testing.T) {
 		}
 	}
 }
+
+func BenchmarkXchg8(b *testing.B) {
+	var x [512]uint8 // give byte its own cache line
+	sink = &x
+	for i := 0; i < b.N; i++ {
+		atomic.Xchg8(&x[255], uint8(i))
+	}
+}
+
+func BenchmarkXchg8Parallel(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.Xchg8(&x[255], i)
+			i++
+		}
+	})
+}