cat >umod-amd64.s <<'EOF'
# tu_int __umodti3(tu_int x, tu_int y)
# x is rsi:rdi, y is rcx:rdx, return result is rdx:rax.
.globl __umodti3
__umodti3:
	# specialized to u128 % u64, so verify that
	test %rcx,%rcx
	jne 1f

	# save divisor
	movq %rdx, %r8

	# reduce top 64 bits mod divisor
	movq %rsi, %rax
	xorl %edx, %edx
	divq %r8

	# reduce full 128-bit mod divisor

	// DX: ULONG64 EstablisherFrame
	// R8: PCONTEXT ContextRecord
	// R9: PDISPATCHER_CONTEXT DispatcherContext
	// Switch from the host ABI to the Go ABI.
	PUSH_REGS_HOST_TO_ABI0()

	get_tls(AX)
	CMPQ	AX, $0
	JNE	2(PC)
	// This shouldn't happen, sehtramp is only attached to functions
	// called from Go, and exception handlers are only called from
	// the thread that threw the exception.
	INT	$3