diff --git a/sys/compat/ndis/winx64_wrap.S b/sys/compat/ndis/winx64_wrap.S
index a94cd927ede9..1f9fc20b691b 100644
--- a/sys/compat/ndis/winx64_wrap.S
+++ b/sys/compat/ndis/winx64_wrap.S
@@ -91,42 +91,29 @@
 	.globl x86_64_wrap_end
 
 ENTRY(x86_64_wrap)
-	subq	$96,%rsp
-	mov     %rsi,96-8(%rsp)	
-	mov     %rdi,96-16(%rsp)
-	mov     %rcx,%rdi
-	mov     %rdx,%rsi
-	mov     %r8,%rdx
-	mov     %r9,%rcx
-	mov	96+40(%rsp),%r8
-	mov	96+48(%rsp),%r9
-	mov	96+56(%rsp),%rax
-	mov	%rax,(%rsp)
-	mov	96+64(%rsp),%rax
-	mov	%rax,8(%rsp)
-	mov	96+72(%rsp),%rax
-	mov	%rax,16(%rsp)
-	mov	96+80(%rsp),%rax
-	mov	%rax,24(%rsp)
-	mov	96+88(%rsp),%rax
-	mov	%rax,32(%rsp)
-	mov	96+96(%rsp),%rax
-	mov	%rax,40(%rsp)
-	mov	96+104(%rsp),%rax
-	mov	%rax,48(%rsp)
-	mov	96+112(%rsp),%rax
-	mov	%rax,56(%rsp)
-	mov	96+120(%rsp),%rax
-	mov	%rax,64(%rsp)
-	mov	96+128(%rsp),%rax
-	mov	%rax,72(%rsp)
-	xor	%rax,%rax
+	subq	$96,%rsp	# allocate space on stack
+	mov     %rsi,96-8(%rsp)	# save %rsi
+	mov     %rdi,96-16(%rsp)# save %rdi
+	mov	%rcx,%r10	# temporarily save %rcx in scratch
+	mov	%rsp,%rsi
+	add	$96+56,%rsi	# source == old stack top (stack+56)
+	mov	%rsp,%rdi	# destination == new stack top
+	mov	$10,%rcx	# count == 10 quadwords
+	rep
+	movsq			# copy old stack contents to new location
+	mov	%r10,%rdi	# set up arg0 (%rcx -> %rdi)
+	mov     %rdx,%rsi	# set up arg1 (%rdx -> %rsi)
+	mov     %r8,%rdx	# set up arg2 (%r8 -> %rdx)
+	mov     %r9,%rcx	# set up arg3 (%r9 -> %rcx)
+	mov	96+40(%rsp),%r8	# set up arg4 (stack+40 -> %r8)
+	mov	96+48(%rsp),%r9	# set up arg5 (stack+48 -> %r9)
+	xor	%rax,%rax	# clear return value
 x86_64_wrap_call:
 	mov	$0xFF00FF00FF00FF00,%r11
-	callq	*%r11
-	mov	96-16(%rsp),%rdi
-	mov	96-8(%rsp),%rsi
-	addq	$96,%rsp
+	callq	*%r11		# call routine
+	mov	96-16(%rsp),%rdi# restore %rdi
+	mov	96-8(%rsp),%rsi	# restore %rsi
+	addq	$96,%rsp	# delete space on stack
 	ret
 x86_64_wrap_end: