extern asm_sse
extern asm_nq
extern asm_zi
extern asm_z2
extern asm_cos
extern asm_zsn
extern asm_zcs
extern asm_dbl
extern asm_fmt
extern printf

	section	.text
	global rot

rot:
	;fxsave [asm_sse]
	fldpi
	fld qword [asm_z2]
	fmulp st1
	fld qword [asm_zi]
	fmulp st1
	fsincos

	movaps xmm0,[asm_nq]	; old n
	movaps xmm1,xmm0
	mulps xmm1,xmm1		; n^2
	movaps xmm2,xmm1
	shufps xmm2,xmm2,0xe5	; zyxx x
	movaps xmm3,xmm1
	shufps xmm3,xmm3,0xe6	; zyxy y
	movaps xmm4,xmm1
	shufps xmm4,xmm4,0xe7	; zyxz z
	comiss xmm2,xmm3
	jc jl1			; if x<y
	comiss xmm2,xmm4
	jc near cm2		; if x<z
	jmp cm0
jl1:	comiss xmm3,xmm4
	jc near cm2		; if y<z
	jmp cm1

cm0:	movaps xmm5,xmm2
	shufps xmm5,xmm5,0x4e	; xxzy
	addps xmm5,xmm2
	rsqrtps xmm5,xmm5
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x10	; 0x00
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x08	; 00y0
	subps xmm6,xmm7		; 0xY0
	movaps xmm1,xmm5
	shufps xmm1,xmm1,0x00	; x+y
	mulps xmm1,xmm6		; p1
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x40	; x000
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x0c	; 00z0
	subps xmm6,xmm7		; x0Z0
	movaps xmm2,xmm5
	shufps xmm2,xmm2,0x55	; x+z
	mulps xmm2,xmm6		; p2
	jmp jl2

cm1:	movaps xmm5,xmm3
	shufps xmm5,xmm5,0x1b	; yxyz
	addps xmm5,xmm3
	rsqrtps xmm5,xmm5
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x80	; y000
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x30	; 0z00
	subps xmm6,xmm7		; yZ00
	movaps xmm1,xmm5
	shufps xmm1,xmm1,0x00	; x+z
	mulps xmm1,xmm6		; p1
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x08	; 00y0
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x10	; 0x00
	subps xmm6,xmm7		; 0Xy0
	movaps xmm2,xmm5
	shufps xmm2,xmm2,0x55	; x+y
	mulps xmm2,xmm6		; p2
	jmp jl2

cm2:	movaps xmm5,xmm4
	shufps xmm5,xmm5,0x4e	; xzzy
	addps xmm5,xmm4
	rsqrtps xmm5,xmm5
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x0c	; 00z0
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x40	; x000
	subps xmm6,xmm7		; X0z0
	movaps xmm1,xmm5
	shufps xmm1,xmm1,0x55	; x+z
	mulps xmm1,xmm6		; p1
	movaps xmm6,xmm0	; <-->
	shufps xmm6,xmm6,0x30	; 0z00
	movaps xmm7,xmm0
	shufps xmm7,xmm7,0x80	; y000
	subps xmm6,xmm7		; Yz00
	movaps xmm2,xmm5
	shufps xmm2,xmm2,0x00	; y+z
	mulps xmm2,xmm6		; p2


jl2:	movaps xmm3,xmm1
	subps xmm3,xmm2		; p1-p2
	addps xmm2,xmm1		; p1+p2
	movaps xmm4,xmm3
	mulps xmm4,xmm4		; (p1-p2)^2
	movaps xmm5,xmm2
	mulps xmm5,xmm5		; (p1+p2)^2
	movaps xmm6,xmm4
	shufps xmm6,xmm5,0x11	; 0101
	movaps xmm7,xmm4
	shufps xmm7,xmm5,0x22	; 0202
	shufps xmm4,xmm5,0x33	; 0303
	addps xmm4,xmm6
	addps xmm4,xmm7
	rsqrtps xmm4,xmm4
	movaps xmm5,xmm4
	shufps xmm4,xmm4,0x00	; 0000
	shufps xmm5,xmm5,0xaa	; 2222
	mulps xmm3,xmm4		; a1
	mulps xmm2,xmm5		; a2

	fwait
	fstp dword [asm_cos]
	fstp dword [asm_cos+4]

	movlps xmm4,[asm_cos]
	movaps xmm5,xmm4
	shufps xmm5,xmm5,0x00	; cos
	shufps xmm4,xmm4,0x55	; sin
	mulps xmm5,xmm3
	mulps xmm4,xmm2
	addps xmm4,xmm5		; u

	movlps xmm2,[asm_zsn]
	movaps xmm3,xmm2
	shufps xmm2,xmm2,0x00	; sin
	shufps xmm3,xmm3,0x55	; cos
	mulps xmm4,xmm2
	mulps xmm0,xmm3
	addps xmm0,xmm4		; new n
	movaps xmm1,xmm0
	mulps xmm1,xmm1
	shufps xmm1,xmm1,0x39	; rotate
	movss xmm2,xmm1
	shufps xmm1,xmm1,0x39	; rotate
	movss xmm3,xmm1
	shufps xmm1,xmm1,0x39	; rotate
	movss xmm4,xmm1
	addss xmm2,xmm3
	addss xmm2,xmm4
	rsqrtss xmm2,xmm2
	shufps xmm2,xmm2,0x00
	mulps xmm0,xmm2		; new n
jl3:	movaps [asm_nq],xmm0

	;fxrstor [asm_sse]
	ret

        fld dword [asm_cos]
        fstp qword [asm_dbl]
        push dword [asm_dbl+4]
        push dword [asm_dbl]
        push dword [asm_fmt]
        call printf
        add esp, 12