runtime: change osyield to use Windows SwitchToThread

It appears that windows osyield is just 15ms sleep on my computer
(see benchmarks below). Replace NtWaitForSingleObject in osyield
with SwitchToThread (as suggested by Dmitry).

Also add issue #14790 related benchmarks, so we can track perfomance
changes in CL 20834 and CL 20835 and beyond.

Update #14790

benchmark                             old ns/op     new ns/op     delta
BenchmarkChanToSyscallPing1ms         1953200       1953000       -0.01%
BenchmarkChanToSyscallPing15ms        31562904      31248400      -1.00%
BenchmarkSyscallToSyscallPing1ms      5247          4202          -19.92%
BenchmarkSyscallToSyscallPing15ms     5260          4374          -16.84%
BenchmarkChanToChanPing1ms            474           494           +4.22%
BenchmarkChanToChanPing15ms           468           489           +4.49%
BenchmarkOsYield1ms                   980018        75.5          -99.99%
BenchmarkOsYield15ms                  15625200      75.8          -100.00%

Change-Id: I1b4cc7caca784e2548ee3c846ca07ef152ebedce
Reviewed-on: https://go-review.googlesource.com/21294
Run-TryBot: Alex Brainman <[email protected]>
Reviewed-by: Dmitry Vyukov <[email protected]>
Run-TryBot: Dmitry Vyukov <[email protected]>
TryBot-Result: Gobot Gobot <[email protected]>

Author: alexbrainman

src/runtime/export_windows_test.go

@@ -9,6 +9,7 @@ package runtime
 import "unsafe"
 
 var TestingWER = &testingWER
+var OsYield = osyield
 
 func NumberOfProcessors() int32 {
 	var info systeminfo

src/runtime/os1_windows.go

@@ -41,6 +41,7 @@ import (
 //go:cgo_import_dynamic runtime._SetUnhandledExceptionFilter SetUnhandledExceptionFilter%1 "kernel32.dll"
 //go:cgo_import_dynamic runtime._SetWaitableTimer SetWaitableTimer%6 "kernel32.dll"
 //go:cgo_import_dynamic runtime._SuspendThread SuspendThread%1 "kernel32.dll"
+//go:cgo_import_dynamic runtime._SwitchToThread SwitchToThread%0 "kernel32.dll"
 //go:cgo_import_dynamic runtime._VirtualAlloc VirtualAlloc%4 "kernel32.dll"
 //go:cgo_import_dynamic runtime._VirtualFree VirtualFree%3 "kernel32.dll"
 //go:cgo_import_dynamic runtime._WSAGetOverlappedResult WSAGetOverlappedResult%5 "ws2_32.dll"
@@ -84,6 +85,7 @@ var (
 	_SetUnhandledExceptionFilter,
 	_SetWaitableTimer,
 	_SuspendThread,
+	_SwitchToThread,
 	_VirtualAlloc,
 	_VirtualFree,
 	_WSAGetOverlappedResult,
@@ -189,6 +191,8 @@ var useLoadLibraryEx bool
 
 func osinit() {
 	asmstdcallAddr = unsafe.Pointer(funcPC(asmstdcall))
+	usleep2Addr = unsafe.Pointer(funcPC(usleep2))
+	switchtothreadAddr = unsafe.Pointer(funcPC(switchtothread))
 
 	setBadSignalMsg()
 
@@ -586,17 +590,22 @@ func stdcall7(fn stdFunction, a0, a1, a2, a3, a4, a5, a6 uintptr) uintptr {
 }
 
 // in sys_windows_386.s and sys_windows_amd64.s
-func usleep1(usec uint32)
+func onosstack(fn unsafe.Pointer, arg uint32)
+func usleep2(usec uint32)
+func switchtothread()
+
+var usleep2Addr unsafe.Pointer
+var switchtothreadAddr unsafe.Pointer
 
 //go:nosplit
 func osyield() {
-	usleep1(1)
+	onosstack(switchtothreadAddr, 0)
 }
 
 //go:nosplit
 func usleep(us uint32) {
 	// Have 1us units; want 100ns units.
-	usleep1(10 * us)
+	onosstack(usleep2Addr, 10*us)
 }
 
 func ctrlhandler1(_type uint32) uint32 {

src/runtime/sys_windows_386.s

@@ -358,10 +358,11 @@ TEXT runtime·setldt(SB),NOSPLIT,$0
 	MOVL	CX, 0x14(FS)
 	RET
 
-// Sleep duration is in 100ns units.
-TEXT runtime·usleep1(SB),NOSPLIT,$0
-	MOVL	usec+0(FP), BX
-	MOVL	$runtime·usleep2(SB), AX // to hide from 8l
+// onosstack calls fn on OS stack.
+// func onosstack(fn unsafe.Pointer, arg uint32)
+TEXT runtime·onosstack(SB),NOSPLIT,$0
+	MOVL	fn+0(FP), AX		// to hide from 8l
+	MOVL	arg+4(FP), BX
 
 	// Execute call on m->g0 stack, in case we are not actually
 	// calling a system call wrapper, like when running under WINE.
@@ -423,6 +424,14 @@ TEXT runtime·usleep2(SB),NOSPLIT,$20
 	MOVL	BP, SP
 	RET
 
+// Runs on OS stack.
+TEXT runtime·switchtothread(SB),NOSPLIT,$0
+	MOVL	SP, BP
+	MOVL	runtime·_SwitchToThread(SB), AX
+	CALL	AX
+	MOVL	BP, SP
+	RET
+
 // func now() (sec int64, nsec int32)
 TEXT time·now(SB),NOSPLIT,$8-12
 	CALL	runtime·unixnano(SB)

src/runtime/sys_windows_amd64.s

@@ -381,10 +381,10 @@ TEXT runtime·settls(SB),NOSPLIT,$0
 	MOVQ	DI, 0x28(GS)
 	RET
 
-// Sleep duration is in 100ns units.
-TEXT runtime·usleep1(SB),NOSPLIT,$0
-	MOVL	usec+0(FP), BX
-	MOVQ	$runtime·usleep2(SB), AX // to hide from 6l
+// func onosstack(fn unsafe.Pointer, arg uint32)
+TEXT runtime·onosstack(SB),NOSPLIT,$0
+	MOVQ	fn+0(FP), AX		// to hide from 6l
+	MOVL	arg+8(FP), BX
 
 	// Execute call on m->g0 stack, in case we are not actually
 	// calling a system call wrapper, like when running under WINE.
@@ -445,6 +445,18 @@ TEXT runtime·usleep2(SB),NOSPLIT,$48
 	MOVQ	40(SP), SP
 	RET
 
+// Runs on OS stack.
+TEXT runtime·switchtothread(SB),NOSPLIT,$0
+	MOVQ	SP, AX
+	ANDQ	$~15, SP	// alignment as per Windows requirement
+	SUBQ	$(48), SP	// room for SP and 4 args as per Windows requirement
+				// plus one extra word to keep stack 16 bytes aligned
+	MOVQ	AX, 32(SP)
+	MOVQ	runtime·_SwitchToThread(SB), AX
+	CALL	AX
+	MOVQ	32(SP), SP
+	RET
+
 // func now() (sec int64, nsec int32)
 TEXT time·now(SB),NOSPLIT,$8-12
 	CALL	runtime·unixnano(SB)

src/runtime/syscall_windows_test.go

@@ -864,3 +864,147 @@ func TestLoadLibraryEx(t *testing.T) {
 	t.Skipf("LoadLibraryEx not usable, but not expected. (LoadLibraryEx=%v; flags=%v)",
 		have, flags)
 }
+
+var (
+	modwinmm    = syscall.NewLazyDLL("winmm.dll")
+	modkernel32 = syscall.NewLazyDLL("kernel32.dll")
+
+	proctimeBeginPeriod = modwinmm.NewProc("timeBeginPeriod")
+	proctimeEndPeriod   = modwinmm.NewProc("timeEndPeriod")
+
+	procCreateEvent = modkernel32.NewProc("CreateEventW")
+	procSetEvent    = modkernel32.NewProc("SetEvent")
+)
+
+func timeBeginPeriod(period uint32) {
+	syscall.Syscall(proctimeBeginPeriod.Addr(), 1, uintptr(period), 0, 0)
+}
+
+func timeEndPeriod(period uint32) {
+	syscall.Syscall(proctimeEndPeriod.Addr(), 1, uintptr(period), 0, 0)
+}
+
+func createEvent() (syscall.Handle, error) {
+	r0, _, e0 := syscall.Syscall6(procCreateEvent.Addr(), 4, 0, 0, 0, 0, 0, 0)
+	if r0 == 0 {
+		return 0, syscall.Errno(e0)
+	}
+	return syscall.Handle(r0), nil
+}
+
+func setEvent(h syscall.Handle) error {
+	r0, _, e0 := syscall.Syscall(procSetEvent.Addr(), 1, uintptr(h), 0, 0)
+	if r0 == 0 {
+		return syscall.Errno(e0)
+	}
+	return nil
+}
+
+func benchChanToSyscallPing(b *testing.B) {
+	ch := make(chan int)
+	event, err := createEvent()
+	if err != nil {
+		b.Fatal(err)
+	}
+	go func() {
+		for i := 0; i < b.N; i++ {
+			syscall.WaitForSingleObject(event, syscall.INFINITE)
+			ch <- 1
+		}
+	}()
+	for i := 0; i < b.N; i++ {
+		err := setEvent(event)
+		if err != nil {
+			b.Fatal(err)
+		}
+		<-ch
+	}
+}
+
+func BenchmarkChanToSyscallPing1ms(b *testing.B) {
+	timeBeginPeriod(1)
+	benchChanToSyscallPing(b)
+	timeEndPeriod(1)
+}
+
+func BenchmarkChanToSyscallPing15ms(b *testing.B) {
+	benchChanToSyscallPing(b)
+}
+
+func benchSyscallToSyscallPing(b *testing.B) {
+	event1, err := createEvent()
+	if err != nil {
+		b.Fatal(err)
+	}
+	event2, err := createEvent()
+	if err != nil {
+		b.Fatal(err)
+	}
+	go func() {
+		for i := 0; i < b.N; i++ {
+			syscall.WaitForSingleObject(event1, syscall.INFINITE)
+			err := setEvent(event2)
+			if err != nil {
+				b.Fatal(err)
+			}
+		}
+	}()
+	for i := 0; i < b.N; i++ {
+		err := setEvent(event1)
+		if err != nil {
+			b.Fatal(err)
+		}
+		syscall.WaitForSingleObject(event2, syscall.INFINITE)
+	}
+}
+
+func BenchmarkSyscallToSyscallPing1ms(b *testing.B) {
+	timeBeginPeriod(1)
+	benchSyscallToSyscallPing(b)
+	timeEndPeriod(1)
+}
+
+func BenchmarkSyscallToSyscallPing15ms(b *testing.B) {
+	benchSyscallToSyscallPing(b)
+}
+
+func benchChanToChanPing(b *testing.B) {
+	ch1 := make(chan int)
+	ch2 := make(chan int)
+	go func() {
+		for i := 0; i < b.N; i++ {
+			<-ch1
+			ch2 <- 1
+		}
+	}()
+	for i := 0; i < b.N; i++ {
+		ch1 <- 1
+		<-ch2
+	}
+}
+
+func BenchmarkChanToChanPing1ms(b *testing.B) {
+	timeBeginPeriod(1)
+	benchChanToChanPing(b)
+	timeEndPeriod(1)
+}
+
+func BenchmarkChanToChanPing15ms(b *testing.B) {
+	benchChanToChanPing(b)
+}
+
+func benchOsYield(b *testing.B) {
+	for i := 0; i < b.N; i++ {
+		runtime.OsYield()
+	}
+}
+
+func BenchmarkOsYield1ms(b *testing.B) {
+	timeBeginPeriod(1)
+	benchOsYield(b)
+	timeEndPeriod(1)
+}
+
+func BenchmarkOsYield15ms(b *testing.B) {
+	benchOsYield(b)
+}