runtime: convert exit to use pthread library on Darwin

Now we no longer need to mess with TLS on Darwin 386/amd64, we always
rely on the pthread library to set it up. We now just use one entry
in the TLS for the G.
Return from mstart to let the pthread library clean up the OS thread.

Change-Id: Iccf58049d545515d9b1d090b161f420e40ffd244
Reviewed-on: https://go-review.googlesource.com/110215Reviewed-by: Ian Lance Taylor <iant@golang.org>

src/runtime/asm_386.s

@@ -228,6 +228,10 @@ needtls:
 	// skip runtime·ldt0setup(SB) and tls test on Plan 9 in all cases
 	JMP	ok
 #endif
+#ifdef GOOS_darwin
+	// skip runtime·ldt0setup(SB) on Darwin
+	JMP	ok
+#endif
 
 	// set up %gs
 	CALL	runtime·ldt0setup(SB)

src/runtime/asm_amd64.s

@@ -188,6 +188,10 @@ needtls:
 	// skip TLS setup on Solaris
 	JMP ok
 #endif
+#ifdef GOOS_darwin
+	// skip TLS setup on Darwin
+	JMP ok
+#endif
 
 	LEAQ	runtime·m0+m_tls(SB), DI
 	CALL	runtime·settls(SB)

src/runtime/os_darwin.go

@@ -135,11 +135,14 @@ func newosproc(mp *m) {
 		exit(1)
 	}
 
-	// Set the stack we want to use.
-	if pthread_attr_setstack(&attr, unsafe.Pointer(mp.g0.stack.lo), mp.g0.stack.hi-mp.g0.stack.lo) != 0 {
+	// Set the stack size we want to use.  64KB for now.
+	// TODO: just use OS default size?
+	const stackSize = 1 << 16
+	if pthread_attr_setstacksize(&attr, stackSize) != 0 {
 		write(2, unsafe.Pointer(&failthreadcreate[0]), int32(len(failthreadcreate)))
 		exit(1)
 	}
+	//mSysStatInc(&memstats.stacks_sys, stackSize) //TODO: do this?
 
 	// Tell the pthread library we won't join with this thread.
 	if pthread_attr_setdetachstate(&attr, _PTHREAD_CREATE_DETACHED) != 0 {
@@ -169,12 +172,6 @@ func mstart_stub()
 //
 //go:nosplit
 func newosproc0(stacksize uintptr, fn uintptr) {
-	stack := sysAlloc(stacksize, &memstats.stacks_sys)
-	if stack == nil {
-		write(2, unsafe.Pointer(&failallocatestack[0]), int32(len(failallocatestack)))
-		exit(1)
-	}
-
 	// Initialize an attribute object.
 	var attr pthreadattr
 	var err int32
@@ -185,10 +182,11 @@ func newosproc0(stacksize uintptr, fn uintptr) {
 	}
 
 	// Set the stack we want to use.
-	if pthread_attr_setstack_trampoline(&attr, stack, stacksize) != 0 {
+	if pthread_attr_setstacksize_trampoline(&attr, stacksize) != 0 {
 		write(2, unsafe.Pointer(&failthreadcreate[0]), int32(len(failthreadcreate)))
 		exit(1)
 	}
+	mSysStatInc(&memstats.stacks_sys, stacksize)
 
 	// Tell the pthread library we won't join with this thread.
 	if pthread_attr_setdetachstate_trampoline(&attr, _PTHREAD_CREATE_DETACHED) != 0 {

src/runtime/proc.go

@@ -1195,8 +1195,8 @@ func mstart() {
 	mstart1()
 
 	// Exit this thread.
-	if GOOS == "windows" || GOOS == "solaris" || GOOS == "plan9" {
-		// Window, Solaris and Plan 9 always system-allocate
+	if GOOS == "windows" || GOOS == "solaris" || GOOS == "plan9" || (GOOS == "darwin" && (GOARCH == "amd64" || GOARCH == "386")) {
+		// Window, Solaris, Darwin and Plan 9 always system-allocate
 		// the stack, but put it in _g_.stack before mstart,
 		// so the logic above hasn't set osStack yet.
 		osStack = true
@@ -1517,9 +1517,9 @@ func allocm(_p_ *p, fn func()) *m {
 	mp.mstartfn = fn
 	mcommoninit(mp)
 
-	// In case of cgo or Solaris, pthread_create will make us a stack.
+	// In case of cgo or Solaris or Darwin, pthread_create will make us a stack.
 	// Windows and Plan 9 will layout sched stack on OS stack.
-	if iscgo || GOOS == "solaris" || GOOS == "windows" || GOOS == "plan9" {
+	if iscgo || GOOS == "solaris" || GOOS == "windows" || GOOS == "plan9" || (GOOS == "darwin" && (GOARCH == "386" || GOARCH == "amd64")) {
 		mp.g0 = malg(-1)
 	} else {
 		mp.g0 = malg(8192 * sys.StackGuardMultiplier)

src/runtime/sys_darwin.go

@@ -21,15 +21,15 @@ func pthread_attr_init(attr *pthreadattr) (errno int32) {
 func pthread_attr_init_trampoline(attr *pthreadattr) int32
 
 //go:nowritebarrier
-func pthread_attr_setstack(attr *pthreadattr, addr unsafe.Pointer, size uintptr) (errno int32) {
+func pthread_attr_setstacksize(attr *pthreadattr, size uintptr) (errno int32) {
 	systemstack(func() {
-		errno = pthread_attr_setstack_trampoline(attr, addr, size)
+		errno = pthread_attr_setstacksize_trampoline(attr, size)
 	})
 	return
 }
 
 //go:noescape
-func pthread_attr_setstack_trampoline(attr *pthreadattr, addr unsafe.Pointer, size uintptr) int32
+func pthread_attr_setstacksize_trampoline(attr *pthreadattr, size uintptr) int32
 
 //go:nowritebarrier
 func pthread_attr_setdetachstate(attr *pthreadattr, state int) (errno int32) {
@@ -57,9 +57,10 @@ func pthread_create_trampoline(t *pthread, attr *pthreadattr, start uintptr, arg
 // in a system library, with the libc_ prefix missing.
 
 //go:cgo_import_dynamic libc_pthread_attr_init pthread_attr_init "/usr/lib/libSystem.B.dylib"
-//go:cgo_import_dynamic libc_pthread_attr_setstack pthread_attr_setstack "/usr/lib/libSystem.B.dylib"
+//go:cgo_import_dynamic libc_pthread_attr_setstacksize pthread_attr_setstacksize "/usr/lib/libSystem.B.dylib"
 //go:cgo_import_dynamic libc_pthread_attr_setdetachstate pthread_attr_setdetachstate "/usr/lib/libSystem.B.dylib"
 //go:cgo_import_dynamic libc_pthread_create pthread_create "/usr/lib/libSystem.B.dylib"
+//go:cgo_import_dynamic libc_exit exit "/usr/lib/libSystem.B.dylib"
 
 // Magic incantation to get libSystem actually dynamically linked.
 // TODO: Why does the code require this?  See cmd/compile/internal/ld/go.go:210

src/runtime/sys_darwin_386.s

@@ -11,46 +11,23 @@
 #include "textflag.h"
 
 // Exit the entire program (like C exit)
-TEXT runtime·exit(SB),NOSPLIT,$0
-	MOVL	$1, AX
-	INT	$0x80
-	MOVL	$0xf1, 0xf1  // crash
-	RET
-
-// Exit this OS thread (like pthread_exit, which eventually
-// calls __bsdthread_terminate).
-TEXT exit1<>(SB),NOSPLIT,$16-0
-	// __bsdthread_terminate takes 4 word-size arguments.
-	// Set them all to 0. (None are an exit status.)
-	MOVL	$0, 0(SP)
-	MOVL	$0, 4(SP)
-	MOVL	$0, 8(SP)
-	MOVL	$0, 12(SP)
-	MOVL	$361, AX
-	INT	$0x80
-	JAE 2(PC)
+TEXT runtime·exit(SB),NOSPLIT,$0-4
+	MOVL	code+0(FP), AX
+	PUSHL	BP
+	MOVL	SP, BP
+	SUBL	$4, SP   // allocate space for callee args
+	ANDL	$~15, SP // align stack
+	MOVL	AX, 0(SP)
+	CALL	libc_exit(SB)
 	MOVL	$0xf1, 0xf1  // crash
+	MOVL	BP, SP
+	POPL	BP
 	RET
 
-GLOBL exitStack<>(SB),RODATA,$(4*4)
-DATA exitStack<>+0x00(SB)/4, $0
-DATA exitStack<>+0x04(SB)/4, $0
-DATA exitStack<>+0x08(SB)/4, $0
-DATA exitStack<>+0x0c(SB)/4, $0
-
-// func exitThread(wait *uint32)
+// Not used on Darwin.
 TEXT runtime·exitThread(SB),NOSPLIT,$0-4
-	MOVL	wait+0(FP), AX
-	// We're done using the stack.
-	MOVL	$0, (AX)
-	// __bsdthread_terminate takes 4 arguments, which it expects
-	// on the stack. They should all be 0, so switch over to a
-	// fake stack of 0s. It won't write to the stack.
-	MOVL	$exitStack<>(SB), SP
-	MOVL	$361, AX	// __bsdthread_terminate
-	INT	$0x80
 	MOVL	$0xf1, 0xf1  // crash
-	JMP	0(PC)
+	RET
 
 TEXT runtime·open(SB),NOSPLIT,$0
 	MOVL	$5, AX
@@ -437,38 +414,8 @@ TEXT runtime·mach_semaphore_signal_all(SB),NOSPLIT,$0
 	RET
 
 // func setldt(entry int, address int, limit int)
-// entry and limit are ignored.
 TEXT runtime·setldt(SB),NOSPLIT,$32
-	MOVL	address+4(FP), BX	// aka base
-
-	/*
-	 * When linking against the system libraries,
-	 * we use its pthread_create and let it set up %gs
-	 * for us.  When we do that, the private storage
-	 * we get is not at 0(GS) but at 0x18(GS).
-	 * The linker rewrites 0(TLS) into 0x18(GS) for us.
-	 * To accommodate that rewrite, we translate the
-	 * address here so that 0x18(GS) maps to 0(address).
-	 *
-	 * Constant must match the one in cmd/link/internal/ld/sym.go.
-	 */
-	SUBL	$0x18, BX
-
-	/*
-	 * Must set up as USER_CTHREAD segment because
-	 * Darwin forces that value into %gs for signal handlers,
-	 * and if we don't set one up, we'll get a recursive
-	 * fault trying to get into the signal handler.
-	 * Since we have to set one up anyway, it might as
-	 * well be the value we want.  So don't bother with
-	 * i386_set_ldt.
-	 */
-	MOVL	BX, 4(SP)
-	MOVL	$3, AX	// thread_fast_set_cthread_self - machdep call #3
-	INT	$0x82	// sic: 0x82, not 0x80, for machdep call
-
-	XORL	AX, AX
-	MOVW	GS, AX
+	// Nothing to do on Darwin, pthread already set thread-local storage up.
 	RET
 
 TEXT runtime·sysctl(SB),NOSPLIT,$0
@@ -531,11 +478,11 @@ TEXT runtime·mstart_stub(SB),NOSPLIT,$0
 	// Someday the convention will be D is always cleared.
 	CLD
 
-	CALL	runtime·stackcheck(SB) // just in case
 	CALL	runtime·mstart(SB)
 
-	// mstart shouldn't ever return, and if it does, we shouldn't ever join to this thread
-	// to get its return status. But tell pthread everything is ok, just in case.
+	// Go is all done with this OS thread.
+	// Tell pthread everything is ok (we never join with this thread, so
+	// the value here doesn't really matter).
 	XORL	AX, AX
 	RET
 
@@ -565,26 +512,24 @@ TEXT runtime·pthread_attr_init_trampoline(SB),NOSPLIT,$0-8
 	MOVL	AX, ret+4(FP)
 	RET
 
-TEXT runtime·pthread_attr_setstack_trampoline(SB),NOSPLIT,$0-16
+TEXT runtime·pthread_attr_setstacksize_trampoline(SB),NOSPLIT,$0-12
 	MOVL	attr+0(FP), AX
-	MOVL	addr+4(FP), CX
-	MOVL	size+8(FP), DX
+	MOVL	size+4(FP), CX
 
 	PUSHL	BP
 	MOVL	SP, BP
 
-	SUBL	$12, SP
+	SUBL	$8, SP
 	ANDL	$~15, SP
 
 	MOVL	AX, 0(SP)
 	MOVL	CX, 4(SP)
-	MOVL	DX, 8(SP)
-	CALL	libc_pthread_attr_setstack(SB)
+	CALL	libc_pthread_attr_setstacksize(SB)
 
 	MOVL	BP, SP
 	POPL	BP
 
-	MOVL	AX, ret+12(FP)
+	MOVL	AX, ret+8(FP)
 	RET
 
 TEXT runtime·pthread_attr_setdetachstate_trampoline(SB),NOSPLIT,$0-12

src/runtime/sys_darwin_amd64.s

@@ -16,35 +16,22 @@
 #include "textflag.h"
 
 // Exit the entire program (like C exit)
-TEXT runtime·exit(SB),NOSPLIT,$0
+TEXT runtime·exit(SB),NOSPLIT,$0-4
 	MOVL	code+0(FP), DI		// arg 1 exit status
-	MOVL	$(0x2000000+1), AX	// syscall entry
-	SYSCALL
+	PUSHQ	BP
+	MOVQ	SP, BP
+	ANDQ	$~15, SP // align stack
+	CALL	libc_exit(SB)
 	MOVL	$0xf1, 0xf1  // crash
+	MOVQ	BP, SP
+	POPQ	BP
 	RET
 
-// Exit this OS thread (like pthread_exit, which eventually
-// calls __bsdthread_terminate).
-TEXT exit1<>(SB),NOSPLIT,$0
-	// Because of exitThread below, this must not use the stack.
-	// __bsdthread_terminate takes 4 word-size arguments.
-	// Set them all to 0. (None are an exit status.)
-	MOVL	$0, DI
-	MOVL	$0, SI
-	MOVL	$0, DX
-	MOVL	$0, R10
-	MOVL	$(0x2000000+361), AX	// syscall entry
-	SYSCALL
+// Not used on Darwin.
+TEXT runtime·exitThread(SB),NOSPLIT,$0-8
 	MOVL	$0xf1, 0xf1  // crash
 	RET
 
-// func exitThread(wait *uint32)
-TEXT runtime·exitThread(SB),NOSPLIT,$0-8
-	MOVQ	wait+0(FP), AX
-	// We're done using the stack.
-	MOVL	$0, (AX)
-	JMP	exit1<>(SB)
-
 TEXT runtime·open(SB),NOSPLIT,$0
 	MOVQ	name+0(FP), DI		// arg 1 pathname
 	MOVL	mode+8(FP), SI		// arg 2 flags
@@ -490,16 +477,8 @@ TEXT runtime·mach_semaphore_signal_all(SB),NOSPLIT,$0
 	MOVL	AX, ret+8(FP)
 	RET
 
-// set tls base to DI
 TEXT runtime·settls(SB),NOSPLIT,$32
-	/*
-	 * Same as in sys_darwin_386.s, but a different constant.
-	 * Constant must match the one in cmd/link/internal/ld/sym.go.
-	 */
-	SUBQ $0x30, DI
-
-	MOVL	$(0x3000000+3), AX	// thread_fast_set_cthread_self - machdep call #3
-	SYSCALL
+	// Nothing to do on Darwin, pthread already set thread-local storage up.
 	RET
 
 TEXT runtime·sysctl(SB),NOSPLIT,$0
@@ -571,11 +550,11 @@ TEXT runtime·mstart_stub(SB),NOSPLIT,$0
 	// Someday the convention will be D is always cleared.
 	CLD
 
-	CALL	runtime·stackcheck(SB) // just in case
 	CALL	runtime·mstart(SB)
 
-	// mstart shouldn't ever return, and if it does, we shouldn't ever join to this thread
-	// to get its return status. But tell pthread everything is ok, just in case.
+	// Go is all done with this OS thread.
+	// Tell pthread everything is ok (we never join with this thread, so
+	// the value here doesn't really matter).
 	XORL	AX, AX
 	RET
 
@@ -591,17 +570,16 @@ TEXT runtime·pthread_attr_init_trampoline(SB),NOSPLIT,$0-12
 	MOVL	AX, ret+8(FP)
 	RET
 
-TEXT runtime·pthread_attr_setstack_trampoline(SB),NOSPLIT,$0-28
+TEXT runtime·pthread_attr_setstacksize_trampoline(SB),NOSPLIT,$0-20
 	MOVQ	attr+0(FP), DI
-	MOVQ	addr+8(FP), SI
-	MOVQ	size+16(FP), DX
+	MOVQ	size+8(FP), SI
 	PUSHQ	BP
 	MOVQ	SP, BP
 	ANDQ	$~15, SP
-	CALL	libc_pthread_attr_setstack(SB)
+	CALL	libc_pthread_attr_setstacksize(SB)
 	MOVQ	BP, SP
 	POPQ	BP
-	MOVL	AX, ret+24(FP)
+	MOVL	AX, ret+16(FP)
 	RET
 
 TEXT runtime·pthread_attr_setdetachstate_trampoline(SB),NOSPLIT,$0-20