cmd/compile, runtime: make atomic loads/stores sequentially consistent on s390x

The z/Architecture does not guarantee that a load following a store
will not be reordered with that store, unless they access the same
address. Therefore if we want to ensure the sequential consistency
of atomic loads and stores we need to perform serialization
operations after atomic stores.

We do not need to serialize in the runtime when using StoreRel[ease]
and LoadAcq[uire]. The z/Architecture already provides sufficient
ordering guarantees for these operations.

name              old time/op  new time/op  delta
AtomicLoad64-16   0.51ns ± 0%  0.51ns ± 0%     ~     (all equal)
AtomicStore64-16  0.51ns ± 0%  0.60ns ± 9%  +16.47%  (p=0.000 n=17+20)
AtomicLoad-16     0.51ns ± 0%  0.51ns ± 0%     ~     (all equal)
AtomicStore-16    0.51ns ± 0%  0.60ns ± 9%  +16.50%  (p=0.000 n=18+20)

Fixes #32428.

Change-Id: I88d19a4010c46070e4fff4b41587efe4c628d4d9
Reviewed-on: https://go-review.googlesource.com/c/go/+/180439
Run-TryBot: Michael Munday <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>
Reviewed-by: Austin Clements <[email protected]>

Author: mundaym

src/cmd/compile/internal/gc/ssa.go

@@ -3093,7 +3093,7 @@ func init() {
 			s.vars[&memVar] = s.newValue1(ssa.OpSelect1, types.TypeMem, v)
 			return s.newValue1(ssa.OpSelect0, types.Types[TUINT32], v)
 		},
-		sys.PPC64)
+		sys.PPC64, sys.S390X)
 	addF("runtime/internal/atomic", "Loadp",
 		func(s *state, n *Node, args []*ssa.Value) *ssa.Value {
 			v := s.newValue2(ssa.OpAtomicLoadPtr, types.NewTuple(s.f.Config.Types.BytePtr, types.TypeMem), args[0], s.mem())
@@ -3125,7 +3125,7 @@ func init() {
 			s.vars[&memVar] = s.newValue3(ssa.OpAtomicStoreRel32, types.TypeMem, args[0], args[1], s.mem())
 			return nil
 		},
-		sys.PPC64)
+		sys.PPC64, sys.S390X)
 
 	addF("runtime/internal/atomic", "Xchg",
 		func(s *state, n *Node, args []*ssa.Value) *ssa.Value {

src/cmd/compile/internal/s390x/ssa.go

@@ -800,6 +800,8 @@ func ssaGenValue(s *gc.SSAGenState, v *ssa.Value) {
 		bne := s.Prog(s390x.ABNE)
 		bne.To.Type = obj.TYPE_BRANCH
 		gc.Patch(bne, cs)
+	case ssa.OpS390XSYNC:
+		s.Prog(s390x.ASYNC)
 	case ssa.OpClobber:
 		// TODO: implement for clobberdead experiment. Nop is ok for now.
 	default:

src/cmd/compile/internal/ssa/gen/S390X.rules

@@ -139,16 +139,15 @@
 (RoundToEven x) -> (FIDBR [4] x)
 (Round       x) -> (FIDBR [1] x)
 
-// Atomic loads.
-(AtomicLoad8  ptr mem) -> (MOVBZatomicload ptr mem)
-(AtomicLoad32 ptr mem) -> (MOVWZatomicload ptr mem)
-(AtomicLoad64 ptr mem) -> (MOVDatomicload ptr mem)
-(AtomicLoadPtr ptr mem) -> (MOVDatomicload ptr mem)
-
-// Atomic stores.
-(AtomicStore32 ptr val mem) -> (MOVWatomicstore ptr val mem)
-(AtomicStore64 ptr val mem) -> (MOVDatomicstore ptr val mem)
-(AtomicStorePtrNoWB ptr val mem) -> (MOVDatomicstore ptr val mem)
+// Atomic loads and stores.
+// The SYNC instruction (fast-BCR-serialization) prevents store-load
+// reordering. Other sequences of memory operations (load-load,
+// store-store and load-store) are already guaranteed not to be reordered.
+(AtomicLoad(8|32|Acq32|64|Ptr) ptr mem) -> (MOV(BZ|WZ|WZ|D|D)atomicload ptr mem)
+(AtomicStore(32|64|PtrNoWB) ptr val mem) -> (SYNC (MOV(W|D|D)atomicstore ptr val mem))
+
+// Store-release doesn't require store-load ordering.
+(AtomicStoreRel32 ptr val mem) -> (MOVWatomicstore ptr val mem)
 
 // Atomic adds.
 (AtomicAdd32 ptr val mem) -> (AddTupleFirst32 val (LAA ptr val mem))

src/cmd/compile/internal/ssa/gen/S390XOps.go

@@ -187,6 +187,8 @@ func init() {
 		fpstore    = regInfo{inputs: []regMask{ptrspsb, fp, 0}}
 		fpstoreidx = regInfo{inputs: []regMask{ptrsp, ptrsp, fp, 0}}
 
+		sync = regInfo{inputs: []regMask{0}}
+
 		// LoweredAtomicCas may overwrite arg1, so force it to R0 for now.
 		cas = regInfo{inputs: []regMask{ptrsp, r0, gpsp, 0}, outputs: []regMask{gp, 0}, clobbers: r0}
 
@@ -493,6 +495,9 @@ func init() {
 		{name: "FlagGT"}, // CC=2 (greater than)
 		{name: "FlagOV"}, // CC=3 (overflow)
 
+		// Fast-BCR-serialization to ensure store-load ordering.
+		{name: "SYNC", argLength: 1, reg: sync, asm: "SYNC", typ: "Mem"},
+
 		// Atomic loads. These are just normal loads but return <value,memory> tuples
 		// so they can be properly ordered with other loads.
 		// load from arg0+auxint+aux.  arg1=mem.

src/cmd/compile/internal/ssa/opGen.go

@@ -4,6 +4,30 @@
 
 #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 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 {

src/cmd/compile/internal/ssa/rewriteS390X.go

@@ -36,30 +36,17 @@ func LoadAcq(ptr *uint32) uint32 {
 	return *ptr
 }
 
-//go:noinline
-//go:nosplit
-func Store(ptr *uint32, val uint32) {
-	*ptr = val
-}
-
-//go:noinline
-//go:nosplit
-func Store64(ptr *uint64, val uint64) {
-	*ptr = val
-}
+//go:noescape
+func Store(ptr *uint32, val uint32)
 
-//go:notinheap
-type noWB struct{}
+//go:noescape
+func Store64(ptr *uint64, val uint64)
 
 // NO go:noescape annotation; see atomic_pointer.go.
-//go:noinline
-//go:nosplit
-func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer) {
-	*(**noWB)(ptr) = (*noWB)(val)
-}
+func StorepNoWB(ptr unsafe.Pointer, val unsafe.Pointer)
 
-//go:noinline
 //go:nosplit
+//go:noinline
 func StoreRel(ptr *uint32, val uint32) {
 	*ptr = val
 }