diff --git a/src/pkg/runtime/sys_windows_386.s b/src/pkg/runtime/sys_windows_386.s
index 5290f60934..c62715dbd7 100644
--- a/src/pkg/runtime/sys_windows_386.s
+++ b/src/pkg/runtime/sys_windows_386.s
@@ -38,6 +38,46 @@ TEXT runtime·asmstdcall(SB),7,$0
 
 	RET
 
+TEXT	runtime·write(SB),7,$24
+	// write only writes to stderr; ignore fd
+	MOVL	$-12, 0(SP)
+	MOVL	SP, BP
+	CALL	*runtime·GetStdHandle(SB)
+	MOVL	BP, SP
+	
+	MOVL	AX, 0(SP)	// handle
+	MOVL	buf+4(FP), DX // pointer
+	MOVL	DX, 4(SP)
+	MOVL	count+8(FP), DX // count
+	MOVL	DX, 8(SP)
+	LEAL	20(SP), DX  // written count
+	MOVL	$0, 0(DX)
+	MOVL	DX, 12(SP)
+	MOVL	$0, 16(SP) // overlapped
+	CALL	*runtime·WriteFile(SB)
+	MOVL	BP, SI
+	RET
+
+TEXT	runtime·badcallback(SB),7,$24
+	// write only writes to stderr; ignore fd
+	MOVL	$-12, 0(SP)
+	MOVL	SP, BP
+	CALL	*runtime·GetStdHandle(SB)
+	MOVL	BP, SP
+	
+	MOVL	AX, 0(SP)	// handle
+	MOVL	$runtime·badcallbackmsg(SB), DX // pointer
+	MOVL	DX, 4(SP)
+	MOVL	runtime·badcallbacklen(SB), DX // count
+	MOVL	DX, 8(SP)
+	LEAL	20(SP), DX  // written count
+	MOVL	$0, 0(DX)
+	MOVL	DX, 12(SP)
+	MOVL	$0, 16(SP) // overlapped
+	CALL	*runtime·WriteFile(SB)
+	MOVL	BP, SI
+	RET
+
 // faster get/set last error
 TEXT runtime·getlasterror(SB),7,$0
 	MOVL	0x34(FS), AX
diff --git a/src/pkg/runtime/sys_windows_amd64.s b/src/pkg/runtime/sys_windows_amd64.s
index fa8d4221e2..73dc542aac 100644
--- a/src/pkg/runtime/sys_windows_amd64.s
+++ b/src/pkg/runtime/sys_windows_amd64.s
@@ -60,6 +60,49 @@ loadregs:
 
 	RET
 
+TEXT runtime·write(SB),7,$48
+	// write only ever writes to stderr; ignore fd
+	MOVQ	$-12, CX // stderr
+	MOVQ	CX, 0(SP)
+	MOVQ	runtime·GetStdHandle(SB), AX
+	CALL	AX
+
+	MOVQ	AX, CX	// handle
+	MOVQ	CX, 0(SP)
+	MOVQ	buf+8(FP), DX // pointer
+	MOVQ	DX, 8(SP)
+	MOVL	count+16(FP), R8 // count
+	MOVQ	R8, 16(SP)
+	LEAQ	40(SP), R9  // written count
+	MOVQ	$0, 0(R9)
+	MOVQ	R9, 24(SP)
+	MOVQ	$0, 32(SP)	// overlapped
+	MOVQ	runtime·WriteFile(SB), AX
+	CALL	AX
+	
+	RET
+
+TEXT runtime·badcallback(SB),7,$48
+	MOVQ	$-12, CX // stderr
+	MOVQ	CX, 0(SP)
+	MOVQ	runtime·GetStdHandle(SB), AX
+	CALL	AX
+
+	MOVQ	AX, CX	// handle
+	MOVQ	CX, 0(SP)
+	MOVQ	$runtime·badcallbackmsg(SB), DX // pointer
+	MOVQ	DX, 8(SP)
+	MOVL	$runtime·badcallbacklen(SB), R8 // count
+	MOVQ	R8, 16(SP)
+	LEAQ	40(SP), R9  // written count
+	MOVQ	$0, 0(R9)
+	MOVQ	R9, 24(SP)
+	MOVQ	$0, 32(SP)	// overlapped
+	MOVQ	runtime·WriteFile(SB), AX
+	CALL	AX
+	
+	RET
+
 // faster get/set last error
 TEXT runtime·getlasterror(SB),7,$0
 	MOVQ	0x30(GS), AX
@@ -207,15 +250,18 @@ TEXT runtime·callbackasm(SB),7,$0
 	MOVQ	R14, 8(SP)
 	MOVQ	R15, 0(SP)
 
+	// prepare call stack.  use SUBQ to hide from stack frame checks
 	// cgocallback(void (*fn)(void*), void *frame, uintptr framesize)
-	PUSHQ	DX    // uintptr framesize
-	PUSHQ	CX    // void *frame
-	PUSHQ	AX    // void (*fn)(void*)
+	SUBQ	$24, SP
+	MOVQ	DX, 16(SP)	// uintptr framesize
+	MOVQ	CX, 8(SP)   // void *frame
+	MOVQ	AX, 0(SP)    // void (*fn)(void*)
 	CLD
 	CALL  runtime·cgocallback(SB)
-	POPQ	AX
-	POPQ	CX
-	POPQ	DX
+	MOVQ	0(SP), AX
+	MOVQ	8(SP), CX
+	MOVQ	16(SP), DX
+	ADDQ	$24, SP
 
 	// restore registers as required for windows callback
 	// 6l does not allow writing many POPs here issuing a warning "nosplit stack overflow"
diff --git a/src/pkg/runtime/thread_windows.c b/src/pkg/runtime/thread_windows.c
index e75e0c1569..49beba5dc1 100644
--- a/src/pkg/runtime/thread_windows.c
+++ b/src/pkg/runtime/thread_windows.c
@@ -114,27 +114,6 @@ runtime·exit(int32 code)
 	runtime·stdcall(runtime·ExitProcess, 1, (uintptr)code);
 }
 
-int32
-runtime·write(int32 fd, void *buf, int32 n)
-{
-	void *handle;
-	uint32 written;
-
-	written = 0;
-	switch(fd) {
-	case 1:
-		handle = runtime·stdcall(runtime·GetStdHandle, 1, (uintptr)-11);
-		break;
-	case 2:
-		handle = runtime·stdcall(runtime·GetStdHandle, 1, (uintptr)-12);
-		break;
-	default:
-		return -1;
-	}
-	runtime·stdcall(runtime·WriteFile, 5, handle, buf, (uintptr)n, &written, (uintptr)0);
-	return written;
-}
-
 void
 runtime·osyield(void)
 {
@@ -423,21 +402,5 @@ runtime·setprof(bool on)
 	USED(on);
 }
 
-static int8 badcallback[] = "runtime: cgo callback on thread not created by Go.\n";
-
-// This runs on a foreign stack, without an m or a g.  No stack split.
-#pragma textflag 7
-void
-runtime·badcallback(void)
-{
-	uint32 written;
-
-	runtime·stdcall(
-		runtime·WriteFile, 5,
-		runtime·stdcall(runtime·GetStdHandle, 1, (uintptr)-12), // stderr
-		badcallback,
-		(uintptr)(sizeof badcallback - 1),
-		&written,
-		nil
-	);
-}
+int8 runtime·badcallbackmsg[] = "runtime: cgo callback on thread not created by Go.\n";
+int32 runtime·badcallbacklen = sizeof runtime·badcallbackmsg - 1;