; callf1_64.asm  a basic structure for a subroutine to be called from "C"
;
; Parameters:   double *L
; Result: L[0]=L[0]+3.0  L[1]=L[1]+4.0

        global callf1_64	; linker must know name of subroutine

        extern	printf		; the C function, to be called for demo

        SECTION .data		; Data section, initialized variables
fmt1:	db "rdi=%ld, L[0]=%e", 10, 0	; The printf format, "\n",'0'
fmt2:	db "rdi=%ld, L[1]=%e", 10, 0	; The printf format, "\n",'0'
a3:	dq	3.0		; 64-bit variable a initialized to 3.0
a4:	dq	4.0		; 64-bit variable b initializes to 4.0

	SECTION .bss
a:	resq	1		; temp for saving address

	SECTION .text           ; Code section.

callf1_64:			; name must appear as a nasm label
        push	rbp		; save rbp
;	mov	rbp, rsp	; rbp is callers stack
        push	rdx		; save registers, for debug print
        push	rdi
	push	rsi

	mov	rax,rdi		; first, only, in parameter
	mov	[a],rdi		; save for later use

;	mov	rdi,fmt1	; format for printf debug, demo
;	mov	rsi,rax         ; first parameter for printf
;       movq    xmm0, qword [rax] ; second parameter for printf
;	mov	rax,1		; one xmm registers
;	call    printf		; Call C function	

;	mov	rax,[a]		; first, only, in parameter, demo
;	mov	rdi,fmt2	; format for printf
;	mov	rsi,rax         ; first parameter for printf
;	movq    xmm0, qword [rax+8] ; second parameter for printf
;	mov	rax,1		; one xmm registers
;	call    printf		; Call C function	

	mov	rax,[a]		; add 3.0 to L[0]
	fld	qword [rax] 	; load L[0] (pushed on flt pt stack, st0)
	fadd	qword [a3]	; floating add 3.0 (to st0)
	fstp	qword [rax]	; store into L[0] (pop flt pt stack)

	fld	qword [rax+8] 	; load L[1] (pushed on flt pt stack, st0)
	fadd	qword [a4]	; floating add 4.0 (to st0)
	fstp	qword [rax+8]	; store into L[1] (pop flt pt stack)

        pop	rsi		; restore registers
	pop	rdi		; in reverse order
        pop	rdx
;	mov	rsp,rbp		
        pop	rbp		; restore callers stack frame
        ret			; return