mult_f1?:
; o: r0=(r3*f(r0,r2))<<16
; i: r0, r2, r3

cmp/hs  r0, r2
bt      loc_52D8        ; r2=r2-r0
mov.l   r0, @-r15
sub     r2, r0          ; r0=r0-r2
sts     macl, r1
mul.l   r3, r0          ; r0=r0*r3
sts     macl, r0
lds     r1, macl
swap.w  r0, r0
mov.l   @r15+, r2
extu.w  r0, r0
sub     r0, r2          ; r2=r2-r0
rts
mov     r2, r0

loc_52D8:               ; r2=r2-r0
sub     r0, r2
sts     macl, r1
mul.l   r3, r2
sts     macl, r2        ; r2=r2*r3
lds     r1, macl
swap.w  r2, r2
extu.w  r2, r2
rts
add     r2, r0