; Multiply the 128-bit unsigned numbers in registers RCX:RDX (Y1:Y0) and RSI:RDI
; (Z1:Z0) and return the 256-bit result in registers RCX:RDX:RSI:RDI (A:B:C:D).
; Algorithm: Peter Norton, "Advanced Assembly Language", 1991, pp. 229-230.

option prologue:none, epilogue:none
.code
umul128	proc
	MOV	R10,RDI	   ; Z0
	MULX	R9,RAX,RSI ; R9:RAX = Z1 * Y0
	MULX	R8,RDI,RDI ; R8:RDI = Z0 * Y0
	MOV	RDX,RCX	   ; Y1
	MULX	RCX,R11,RSI; RCX:R11 = Z1 * Y1
	MULX	RDX,RSI,R10; RDX:RSI = Z0 * Y1
	ADD	RSI,R8	   ; C = High(Z0 * Y0) + Low (Z0 * Y1)
	ADC	RDX,R9	   ; B = High(Z0 * Y1) + High(Z1 * Y0)
	ADC	RCX,0	   ; A = High(Z1 * Y1)
	ADD	RSI,RAX	; C = High(Z0 * Y0) + Low (Z0 * Y1) + Low(Z1 * Y0)
	ADC	RDX,R11	; B = High(Z0 * Y1) + High(Z1 * Y0) + Low(Z1 * Y1)
	ADC	RCX,0
	RET
umul128	endp
end