diff --git a/src/cmd/6l/pass.c b/src/cmd/6l/pass.c
index e5b948a6ed..7a95f7427e 100644
--- a/src/cmd/6l/pass.c
+++ b/src/cmd/6l/pass.c
@@ -638,6 +638,9 @@ dostkoff(void)
 
 			q = P;
 			q1 = P;
+			if((p->from.scale & NOSPLIT) && autoffset >= StackSmall)
+				diag("nosplit func likely to overflow stack");
+
 			if(!(p->from.scale & NOSPLIT)) {
 				if(debug['K']) {
 					// 6l -K means check not only for stack
@@ -792,8 +795,44 @@ dostkoff(void)
 				if(q != P)
 					q->pcond = p;
 			}
-
 			deltasp = autoffset;
+
+			if(debug['K'] > 1 && autoffset) {
+				// 6l -KK means double-check for stack overflow
+				// even after calling morestack and even if the
+				// function is marked as nosplit.
+				p = appendp(p);
+				p->as = AMOVQ;
+				p->from.type = D_INDIR+D_R15;
+				p->from.offset = 0;
+				p->to.type = D_BX;
+
+				p = appendp(p);
+				p->as = ASUBQ;
+				p->from.type = D_CONST;
+				p->from.offset = StackSmall+32;
+				p->to.type = D_BX;
+
+				p = appendp(p);
+				p->as = ACMPQ;
+				p->from.type = D_SP;
+				p->to.type = D_BX;
+
+				p = appendp(p);
+				p->as = AJHI;
+				p->to.type = D_BRANCH;
+				q1 = p;
+
+				p = appendp(p);
+				p->as = AINT;
+				p->from.type = D_CONST;
+				p->from.offset = 3;
+
+				p = appendp(p);
+				p->as = ANOP;
+				q1->pcond = p;
+				q1 = P;
+			}
 		}
 		pcsize = p->mode/8;
 		a = p->from.type;
@@ -844,13 +883,12 @@ dostkoff(void)
 			goto become;
 
 		if(autoffset) {
-			q = p;
+			p->as = AADJSP;
+			p->from.type = D_CONST;
+			p->from.offset = -autoffset;
+
 			p = appendp(p);
 			p->as = ARET;
-
-			q->as = AADJSP;
-			q->from.type = D_CONST;
-			q->from.offset = -autoffset;
 		}
 		continue;
 
diff --git a/src/pkg/runtime/386/asm.s b/src/pkg/runtime/386/asm.s
index 5aa73a6b8f..9df7fb1466 100644
--- a/src/pkg/runtime/386/asm.s
+++ b/src/pkg/runtime/386/asm.s
@@ -300,12 +300,22 @@ TEXT	abort(SB),7,$0
 	INT $0x3
 
 // runcgo(void(*fn)(void*), void *arg)
-// Just call fn(arg), but first align the stack
-// appropriately for the gcc ABI.
+// Call fn(arg) on the scheduler stack,
+// aligned appropriately for the gcc ABI.
 TEXT	runcgo(SB),7,$16
 	MOVL	fn+0(FP), AX
 	MOVL	arg+4(FP), BX
 	MOVL	SP, CX
+
+	// Figure out if we need to switch to m->g0 stack.
+	MOVL	m, DX
+	MOVL	m_g0(DX), SI
+	CMPL	g, SI
+	JEQ	2(PC)
+	MOVL	(m_sched+gobuf_sp)(DX), SP
+
+	// Now on a scheduling stack (a pthread-created stack).
+	SUBL	$16, SP
 	ANDL	$~15, SP	// alignment for gcc ABI
 	MOVL	CX, 4(SP)
 	MOVL	BX, 0(SP)
diff --git a/src/pkg/runtime/amd64/asm.s b/src/pkg/runtime/amd64/asm.s
index 6cb6d5c77a..87bc222e14 100644
--- a/src/pkg/runtime/amd64/asm.s
+++ b/src/pkg/runtime/amd64/asm.s
@@ -272,20 +272,32 @@ TEXT jmpdefer(SB), 7, $0
 	JMP	AX	// but first run the deferred function
 
 // runcgo(void(*fn)(void*), void *arg)
-// Call fn(arg), but align the stack
-// appropriately for the gcc ABI
-// and also save g and m across the call,
+// Call fn(arg) on the scheduler stack,
+// aligned appropriately for the gcc ABI.
+// Save g and m across the call,
 // since the foreign code might reuse them.
 TEXT runcgo(SB),7,$32
+	// Save old registers.
 	MOVQ	fn+0(FP),AX
 	MOVQ	arg+8(FP),DI	// DI = first argument in AMD64 ABI
 	MOVQ	SP, CX
+
+	// Figure out if we need to switch to m->g0 stack.
+	MOVQ	m_g0(m), R8
+	CMPQ	R8, g
+	JEQ	2(PC)
+	MOVQ	(m_sched+gobuf_sp)(m), SP
+
+	// Now on a scheduling stack (a pthread-created stack).
+	SUBQ	$32, SP
 	ANDQ	$~15, SP	// alignment for gcc ABI
 	MOVQ	g, 24(SP)	// save old g, m, SP
 	MOVQ	m, 16(SP)
 	MOVQ	CX, 8(SP)
 	CALL	AX
-	MOVQ	16(SP), m	// restore
+
+	// Restore registers, stack pointer.
+	MOVQ	16(SP), m
 	MOVQ	24(SP), g
 	MOVQ	8(SP), SP
 	RET
diff --git a/src/pkg/runtime/cgocall.c b/src/pkg/runtime/cgocall.c
index a475603957..70382ceee1 100644
--- a/src/pkg/runtime/cgocall.c
+++ b/src/pkg/runtime/cgocall.c
@@ -25,10 +25,7 @@ cgocall(void (*fn)(void*), void *arg)
 	 * foreign code.
 	 */
 	sys·entersyscall();
-	g->cgofn = fn;
-	g->cgoarg = arg;
-	g->status = Gcgocall;
-	gosched();
+	runcgo(fn, arg);
 	sys·exitsyscall();
 	return;
 }
diff --git a/src/pkg/runtime/print.c b/src/pkg/runtime/print.c
index fb2881be55..4a358a8116 100644
--- a/src/pkg/runtime/print.c
+++ b/src/pkg/runtime/print.c
@@ -4,7 +4,7 @@
 
 #include "runtime.h"
 
-static Lock debuglock;
+//static Lock debuglock;
 
 void
 dump(byte *p, int32 n)
@@ -37,7 +37,7 @@ printf(int8 *s, ...)
 	int8 *p, *lp;
 	byte *arg, *narg;
 
-	lock(&debuglock);
+//	lock(&debuglock);
 
 	lp = p = s;
 	arg = (byte*)(&s+1);
@@ -100,7 +100,7 @@ printf(int8 *s, ...)
 	if(p > lp)
 		write(1, lp, p-lp);
 
-	unlock(&debuglock);
+//	unlock(&debuglock);
 }
 
 
diff --git a/src/pkg/runtime/proc.c b/src/pkg/runtime/proc.c
index f6f2bb2b36..4113002ada 100644
--- a/src/pkg/runtime/proc.c
+++ b/src/pkg/runtime/proc.c
@@ -95,7 +95,7 @@ schedinit(void)
 {
 	int32 n;
 	byte *p;
-	
+
 	allm = m;
 
 	mallocinit();
@@ -452,15 +452,6 @@ scheduler(void)
 	lock(&sched);
 	if(gosave(&m->sched) != 0){
 		gp = m->curg;
-		if(gp->status == Gcgocall){
-			// Runtime call into external code (FFI).
-			// When running with FFI, the scheduler stack is a
-			// native pthread stack, so it suffices to switch to the
-			// scheduler stack and make the call.
-			runcgo(gp->cgofn, gp->cgoarg);
-			gp->status = Grunning;
-			gogo(&gp->sched, 1);
-		}
 
 		// Jumped here via gosave/gogo, so didn't
 		// execute lock(&sched) above.
diff --git a/src/pkg/runtime/runtime.h b/src/pkg/runtime/runtime.h
index b44eb929cc..b560d68f98 100644
--- a/src/pkg/runtime/runtime.h
+++ b/src/pkg/runtime/runtime.h
@@ -94,7 +94,6 @@ enum
 	Gwaiting,
 	Gmoribund,
 	Gdead,
-	Gcgocall,
 };
 enum
 {