; exnl4.asm  example  loops iteration
	extern printf
        section .data		; Data section, initialized variables
msg:	db "nasm  exnl4.asm running", 0	; C string needs 0
msg2:	db "exnl4.asm finished", 0	
fmt:    db "%s", 10, 0
fmt1:	db "iloop i=%d", 10, 0		; format needs \n 0
n:	dq 3				; for simple loop
fmtfl:	db "floop x=%e", 10, 0
xlast:	dq 8.0  			; for float loop
xinc:	dq 2.0				; for +2.0
m1:	dq 1.1, 1.2, 1.3, 2.1, 2.2, 2.3 ; row major 2 by 3 matrix
m1name:	db "m1", 0
ncol1:	dq 2
nrow1:	dq 3
m2:	dq 1.1, 2.1, 3.1, 1.2, 2.2, 3.2 ; coloum major 3 by 2 matrix
m2name:	db "m2", 0
ncol2:	dq 3
nrow2:	dq 2
i8:	dq 8
fmtm:	db "%s[%d][%d]=%e", 10, 0
fijk:	db "i=%d, j=%d, k=%d, l=%d", 10, 0
	
	section .bss
i:	resq 1			; reserve space, 1 is one quad word
j:	resq 1
k:	resq 1
l:	resq 1
x:	resq 1			; float type or any type
	
        section .text           ; Code section.
        global main
main:
        push    rbp		; set up stack frame, must be aligned
	
	mov	rdi,fmt         ; address of format, required register rdi
	mov	rsi,msg         ; address of data
	mov	rax,0		; or can be  xor  rax,rax
        call    printf		; Call C function

; simple loop
	mov	rax,0		; for(i=0 ; i<n; i++)
	mov	[i],rax		; i=0
loop:				; use i for what you want

	mov	rdi,fmt1	; just print i
        mov	rsi,[i]
	mov	rax,0
	call	printf

	mov	rax,[i]		; safest to keep variable in memory
	inc	rax		; i++
	mov	[i],rax
	cmp	rax,[n]		; i<n
	jne	loop
	

; float loop
	fld	qword [xinc]	; for(x=2.0 x<xlast x=x+xinc)
	fstp	qword [x]	; x=2.0
floop:				; use x for what you want

	mov	rdi,fmtfl	; just print x
        movq	xmm0, qword [x]
	mov	rax,1
	call	printf

	fld	qword [x]	; safest to keep variable in memory
	fadd	qword [xinc]	; x = x + xinc
	fstp	qword [x]
	fld	qword [xlast]	; pushed into st0, then st1
	fld	qword [x]	; pushed into st0
	fcompp			; compare and pop both
;	jg	floop		; infinite loop
	jl	floop		; quit first time


; for(i=0; i<nrow; i++) { for(j=0; j<ncol; j++) { print m1[i][j] }}
	mov	rax,0
	mov	[i],rax		; i=0
loopi:
	mov rax,[i]
	mov rbx,0
	mov [j],rbx		; j=0
loopj:	
	mov	rax,[i]
	mov	rbx,[j]
	imul	qword [ncol1] 	; i*ncol1
	add	rax, rbx	; i*ncol1+j  [i][j]
	mov	[k],rax
	imul	qword [i8]	; byte address offset
	mov	[l],rax		; save to memory

	mov	rdi,fijk	; debug print
        mov	rsi,[i]
	mov	rdx,[j]
	mov	rcx,[k]
	mov	r8,[l]
	mov	rax,0
	call	printf
	
	mov	rdi,fmtm	; print matrix
	mov	rsi,m1name
	mov	rdx,[i]
	mov	rcx,[j]
	mov	rbx,[l]		   ; set up byte offset
	movq	xmm0,qword[m1+rbx] ; m1[(i*ncol1+j)*8] byte address
	mov	rax,1		   ; number of floats to print
	call	printf

	mov	rbx,[j]
	inc	rbx		; j++
	mov	[j],rbx
	cmp 	rbx,[ncol1]	; j<ncol
	jne	loopj

	mov	rax,[i]
	inc	rax		; i++
	mov	[i],rax
	cmp	rax,[nrow1]	; i<nrow1
	jne	loopi


	mov	rdi,fmt		; exit print
	mov	rsi,msg2
	mov	rax,0
	call	printf
        pop     rbp		; restore stack 
	mov	rax,0		; normal, no error, return value
	ret			; return