2 ; Ullrich von Bassewitz, 2009-08-17
4 ; CC65 runtime: multiplication for ints
7 .export tosumulax, tosmulax
8 .import mul8x16, mul8x16a ; in mul8.s
10 .importzp tmp1, ptr1, ptr4
13 ;---------------------------------------------------------------------------
14 ; 16x16 multiplication routine
20 beq @L3 ; Do 8x16 multiplication if high byte zero
21 stx ptr4+1 ; Save right operand
22 jsr popptr1 ; Get left operand (Y=0, X untouched by popptr1)
24 ; Do ptr4:ptr4+1 * ptr1:ptr1+1 --> AX
27 ldy ptr1+1 ; check if lhs is 8 bit only
28 beq @L4 ; -> we can do 8x16 after swap
30 ldy #16 ; Number of bits
33 ror ptr4 ; Get first bit into carry
39 lda ptr1+1 ; Hi byte of left op
51 lda ptr4 ; Load the result
55 ; High byte of rhs is zero, jump to the 8x16 routine instead
59 ; If the high byte of lhs is zero, swap the operands in ptr1/4 and
60 ; use the 8x16 routine. On entry, A and Y are zero and X has the value
63 @L4: stx ptr1+1 ; Store hi-byte from ptr4
64 ldy ptr1 ; Save right operand (8 bit)
65 ldx ptr4 ; Copy left 16 bit operand to right
67 sty ptr4 ; Copy low 8 bit of right op to left
69 jmp mul8x16a ; There, ptr4+1 will be also cleared