diff --git a/src/pkg/runtime/runtime.h b/src/pkg/runtime/runtime.h
index 831510fd6f0..665f15e9241 100644
--- a/src/pkg/runtime/runtime.h
+++ b/src/pkg/runtime/runtime.h
@@ -52,7 +52,7 @@ typedef	struct	G		G;
 typedef	struct	Gobuf		Gobuf;
 typedef	union	Lock		Lock;
 typedef	struct	M		M;
-typedef struct	P		P;
+typedef	struct	P		P;
 typedef	struct	Mem		Mem;
 typedef	union	Note		Note;
 typedef	struct	Slice		Slice;
@@ -265,7 +265,7 @@ struct	M
 	uintptr	cret;		// return value from C
 	uint64	procid;		// for debuggers, but offset not hard-coded
 	G*	gsignal;	// signal-handling G
-	uint32	tls[8];		// thread-local storage (for 386 extern register)
+	uint64	tls[4];		// thread-local storage (for x86 extern register)
 	G*	curg;		// current running goroutine
 	P*	p;		// attached P for executing Go code (nil if not executing Go code)
 	P*	nextp;
diff --git a/src/pkg/runtime/sys_freebsd_386.s b/src/pkg/runtime/sys_freebsd_386.s
index 4a9d7b40faf..f779c666e70 100644
--- a/src/pkg/runtime/sys_freebsd_386.s
+++ b/src/pkg/runtime/sys_freebsd_386.s
@@ -38,7 +38,12 @@ TEXT runtime·thr_start(SB),7,$0
 	
 	MOVL	AX, m(CX)
 	CALL	runtime·stackcheck(SB)		// smashes AX
-	CALL	runtime·mstart(SB)
+
+	// newosproc left the function we should call in mp->tls[2] for us.
+	get_tls(CX)
+	MOVQ	8(CX), AX
+	CALL	AX
+
 	MOVL	0, AX			// crash (not reached)
 
 // Exit the entire program (like C exit)
diff --git a/src/pkg/runtime/sys_freebsd_amd64.s b/src/pkg/runtime/sys_freebsd_amd64.s
index e8c0899cf79..94f37021c53 100644
--- a/src/pkg/runtime/sys_freebsd_amd64.s
+++ b/src/pkg/runtime/sys_freebsd_amd64.s
@@ -34,12 +34,18 @@ TEXT runtime·thr_start(SB),7,$0
 
 	// set up m, g
 	get_tls(CX)
+	MOVQ	8(CX), AX
 	MOVQ	R13, m(CX)
 	MOVQ	m_g0(R13), DI
 	MOVQ	DI, g(CX)
 
 	CALL runtime·stackcheck(SB)
-	CALL runtime·mstart(SB)
+	
+	// newosproc left the function we should call in mp->tls[2] for us.
+	get_tls(CX)
+	MOVQ	16(CX), AX
+	CALL	AX
+
 	MOVQ 0, AX			// crash (not reached)
 
 // Exit the entire program (like C exit)
diff --git a/src/pkg/runtime/sys_freebsd_arm.s b/src/pkg/runtime/sys_freebsd_arm.s
index b6b007759bd..d9e33393243 100644
--- a/src/pkg/runtime/sys_freebsd_arm.s
+++ b/src/pkg/runtime/sys_freebsd_arm.s
@@ -33,7 +33,11 @@ TEXT runtime·thr_start(SB),7,$0
 	// set up g
 	MOVW m_g0(R9), R10
 	BL runtime·emptyfunc(SB) // fault if stack check is wrong
-	BL runtime·mstart(SB)
+
+	// newosproc left the function we should call in mp->tls[2] for us.
+	MOVW	(m_tls+8)(m), R0
+	BL	(R0)
+
 	MOVW $2, R9  // crash (not reached)
 	MOVW R9, (R9)
 	RET
diff --git a/src/pkg/runtime/sys_windows_386.s b/src/pkg/runtime/sys_windows_386.s
index a4ac7463a36..4cb725e197e 100644
--- a/src/pkg/runtime/sys_windows_386.s
+++ b/src/pkg/runtime/sys_windows_386.s
@@ -260,7 +260,10 @@ TEXT runtime·tstart(SB),7,$0
 
 	CALL	runtime·stackcheck(SB)	// clobbers AX,CX
 
-	CALL	runtime·mstart(SB)
+	// start function is in tls[2]
+	get_tls(CX)
+	MOVL	8(CX), AX
+	CALL	AX
 
 	RET
 
diff --git a/src/pkg/runtime/sys_windows_amd64.s b/src/pkg/runtime/sys_windows_amd64.s
index fe88f3b7544..94e5d79917a 100644
--- a/src/pkg/runtime/sys_windows_amd64.s
+++ b/src/pkg/runtime/sys_windows_amd64.s
@@ -329,7 +329,11 @@ TEXT runtime·tstart_stdcall(SB),7,$0
 	CLD
 
 	CALL	runtime·stackcheck(SB)	// clobbers AX,CX
-	CALL	runtime·mstart(SB)
+
+	// start function is in tls[2]
+	get_tls(CX)
+	MOVQ	16(CX), AX
+	CALL	AX
 
 	XORL	AX, AX			// return 0 == success
 	RET
diff --git a/src/pkg/runtime/thread_darwin.c b/src/pkg/runtime/thread_darwin.c
index 1a13eba1cdb..1400e6e7bc9 100644
--- a/src/pkg/runtime/thread_darwin.c
+++ b/src/pkg/runtime/thread_darwin.c
@@ -95,7 +95,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
 		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	runtime·sigprocmask(SIG_SETMASK, &sigset_all, &oset);
diff --git a/src/pkg/runtime/thread_freebsd.c b/src/pkg/runtime/thread_freebsd.c
index d7758eaafbb..06edc22fe1e 100644
--- a/src/pkg/runtime/thread_freebsd.c
+++ b/src/pkg/runtime/thread_freebsd.c
@@ -82,12 +82,13 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	ThrParam param;
 	Sigset oset;
 
-	USED(fn);	// thr_start assumes fn == mstart
-	USED(gp);	// thr_start assumes gp == mp->g0
+	// thr_start assumes gp == mp->g0
+	if(gp != mp->g0)
+		runtime·throw("invalid newosproc gp");
 
 	if(0){
 		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	runtime·sigprocmask(&sigset_all, &oset);
@@ -103,6 +104,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	param.tls_size = sizeof mp->tls;
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
+	mp->tls[2] = (uintptr)fn;
 
 	runtime·thr_new(&param, sizeof param);
 	runtime·sigprocmask(&oset, nil);
diff --git a/src/pkg/runtime/thread_linux.c b/src/pkg/runtime/thread_linux.c
index 85c3e6b8cf6..3f450580692 100644
--- a/src/pkg/runtime/thread_linux.c
+++ b/src/pkg/runtime/thread_linux.c
@@ -143,7 +143,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
 		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p clone=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, runtime·clone, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, runtime·clone, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	// Disable signals during clone, so that the new thread starts
diff --git a/src/pkg/runtime/thread_netbsd.c b/src/pkg/runtime/thread_netbsd.c
index aba8fea7a2e..b9ec33acd5f 100644
--- a/src/pkg/runtime/thread_netbsd.c
+++ b/src/pkg/runtime/thread_netbsd.c
@@ -153,7 +153,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	if(0) {
 		runtime·printf(
 			"newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
diff --git a/src/pkg/runtime/thread_openbsd.c b/src/pkg/runtime/thread_openbsd.c
index 525dc697e01..f35c3bb4413 100644
--- a/src/pkg/runtime/thread_openbsd.c
+++ b/src/pkg/runtime/thread_openbsd.c
@@ -132,7 +132,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	if(0) {
 		runtime·printf(
 			"newosproc stk=%p m=%p g=%p fn=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	mp->tls[0] = mp->id;	// so 386 asm can find it
diff --git a/src/pkg/runtime/thread_plan9.c b/src/pkg/runtime/thread_plan9.c
index f2169e8e68d..866b1e7806f 100644
--- a/src/pkg/runtime/thread_plan9.c
+++ b/src/pkg/runtime/thread_plan9.c
@@ -228,7 +228,7 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	mp->tls[0] = mp->id;	// so 386 asm can find it
 	if(0){
 		runtime·printf("newosproc stk=%p m=%p g=%p fn=%p rfork=%p id=%d/%d ostk=%p\n",
-			stk, mp, gp, fn, runtime·rfork, mp->id, mp->tls[0], &mp);
+			stk, mp, gp, fn, runtime·rfork, mp->id, (int32)mp->tls[0], &mp);
 	}
 
 	if(runtime·rfork(RFPROC|RFMEM|RFNOWAIT, stk, mp, gp, fn) < 0)
diff --git a/src/pkg/runtime/thread_windows.c b/src/pkg/runtime/thread_windows.c
index 4d95e998707..5ff35b811ed 100644
--- a/src/pkg/runtime/thread_windows.c
+++ b/src/pkg/runtime/thread_windows.c
@@ -192,8 +192,12 @@ runtime·newosproc(M *mp, G *gp, void *stk, void (*fn)(void))
 	void *thandle;
 
 	USED(stk);
-	USED(gp);	// assuming gp = mp->g0
-	USED(fn);	// assuming fn = mstart
+
+	// assume gp == mp->g0
+	if(gp != mp->g0)
+		runtime·throw("invalid newosproc gp");
+
+	mp->tls[2] = (uintptr)fn;
 
 	thandle = runtime·stdcall(runtime·CreateThread, 6,
 		nil, (uintptr)0x20000, runtime·tstart_stdcall, mp,