diff --git a/src/pkg/runtime/amd64/asm.s b/src/pkg/runtime/amd64/asm.s
index 627af66320..52b0a89bcb 100644
--- a/src/pkg/runtime/amd64/asm.s
+++ b/src/pkg/runtime/amd64/asm.s
@@ -9,10 +9,16 @@ TEXT	_rt0_amd64(SB),7,$-8
 	MOVQ	0(DI), AX		// argc
 	LEAQ	8(DI), BX		// argv
 	SUBQ	$(4*8+7), SP		// 2args 2auto
-	ANDQ	$~7, SP
+	ANDQ	$~15, SP
 	MOVQ	AX, 16(SP)
 	MOVQ	BX, 24(SP)
 
+	// if there is an initcgo, call it.
+	MOVQ	initcgo(SB), AX
+	TESTQ	AX, AX
+	JZ	2(PC)
+	CALL	AX
+
 	// set the per-goroutine and per-mach registers
 	LEAQ	m0(SB), m
 	LEAQ	g0(SB), g
@@ -372,4 +378,5 @@ TEXT getcallersp(SB),7,$0
 	MOVQ	sp+0(FP), AX
 	RET
 
+GLOBL initcgo(SB), $8
 GLOBL libcgo_set_scheduler(SB), $8
diff --git a/src/pkg/runtime/linux/amd64/rt0.s b/src/pkg/runtime/linux/amd64/rt0.s
index 6b1b297752..2190b44144 100644
--- a/src/pkg/runtime/linux/amd64/rt0.s
+++ b/src/pkg/runtime/linux/amd64/rt0.s
@@ -5,13 +5,6 @@
 // Darwin and Linux use the same linkage to main
 
 TEXT	_rt0_amd64_linux(SB),7,$-8
-	MOVQ	initcgo(SB), AX
-	TESTQ	AX, AX
-	JZ	2(PC)
-	CALL	AX
-
 	MOVQ	$_rt0_amd64(SB), AX
 	MOVQ	SP, DI
 	JMP	AX
-
-GLOBL initcgo(SB), $8