; prtmat.asm  
;
; compile  nasm -f elf64 -l prtmat.lst  prtmat.asm
; link     gcc -m64 -o prtmat  prtmat.o
; run      ./prtmat
;

        extern printf		; the C function to be called

%macro  prtm 1			; print macro, arg1 is string, name of matrix
	section .data           ; for inside macro
	.fmt	db	"%s[%d][%d]=%e", 10, 0
	.str    db      %1,0	; %1 is macro call first actual parameter
	
	section .text
	mov     rdi, .fmt ; address of format string
	mov     rsi, .str ; string passed to macro
	mov	rdx, [i]
	mov	rcx, [j]
	movq    xmm0, qword [x] ; first floating point in fmt
	mov     rax, 1	; 1 floating point arguments to printf
	call    printf	; Call C function
%endmacro
	
	section .data		; initialized
msg:	db "prtmat.asm runnung", 0
msg2:	db "prtmat.asm finished", 0
msg3:	db "using macro to print", 0
msg4:	db "development with debug print", 0
fmt:    db "%s", 10, 0
fmtbl:	db " ", 10, 0			; for blank line
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				; 8 bytes in qword 4*2
fmtm:	db "%s[%d][%d]=%e", 10, 0
fijkl:	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
lm	resq 1			; address of matrix item
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		; no float to print
        call    printf		; Call C function
	
	mov	rdi,fmt         ; address of format, required register rdi
	mov	rsi,msg3
	mov	rax,0		; no float to print
        call    printf		; Call C function

	mov	rax,0		; first print matrix
	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	rax,[ncol1] 	; i*ncol1
	add	rax, rbx	; i*ncol1+j  [i][j]
	mov	[k],rax
	imul	rax,[i8]	; byte address offset
	mov	[l],rax		; (i*ncol1+j)*8

	mov	rbx,[l]
	movq	xmm0,qword[m1+rbx]
        movq	qword[x],xmm0
	prtm "m1"

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

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

	mov	rdi,fmtbl       ; print blank line
	mov	rax,0		; no float to print
        call    printf		; Call C function
	
	mov	rdi,fmt         ; address of format, required register rdi
	mov	rsi,msg4        ; with debug
	mov	rax,0		; no float to print
        call    printf		; Call C function


	mov	rax,0		; now using macro and added debug print
	mov	[i],rax		; i=0
loopi2:
	mov rax,[i]
	mov rbx,0
	mov [j],rbx		; j=0
loopj2:	
	mov	rax,[i]
	mov	rbx,[j]
	imul	rax,[ncol1] 	; i*ncol1
	add	rax, rbx	; i*ncol1+j  [i][j]
	mov	[k],rax
	imul	rax,[i8]	; byte address offset
	mov	[l],rax		; (i*ncol1+j)*8
	
	mov	rdi,fijkl	; debug print needed to develop indexing
	mov	rsi,[i]
	mov	rdx,[j]
	mov	rcx,[k]		; i*ncol1+j
	mov	r8,[l]		; will go into rbx  qword[m1+rbx]
	mov	rax,0
	call	printf

	
	mov	rdi,fmtm
	mov	rsi,m1name
	mov	rdx,[i]
	mov	rcx,[j]
	mov	rbx,[l]
	movq	xmm0,qword[m1+rbx]
	mov	rax,1
	call	printf

	mov	rbx,[j]
	inc	rbx
	mov	[j],rbx
	cmp 	rbx,[ncol1]	; j<ncol1
	jne	loopj2

	mov	rax,[i]
	inc	rax
	mov	[i],rax
	cmp	rax,[nrow1]	; i<nrow1
	jne	loopi2

	mov	rdi,fmtbl       ; print blank line
	mov	rax,0		; no float to print
        call    printf		; Call C function

	mov	rax,0		; print matrix m2
	mov	[i],rax		; i=0
loopi3:
	mov rax,[i]
	mov rbx,0
	mov [j],rbx		; j=0
loopj3:	
	mov	rax,[i]
	mov	rbx,[j]
	imul	rax,[ncol2] 	; i*ncol2
	add	rax, rbx	; i*ncol2+j  [i][j]
	mov	[k],rax
	imul	rax,[i8]	; byte address offset
	mov	[l],rax		; (i*ncol2+j)*8

	mov	rbx,[l]
	movq	xmm0,qword[m1+rbx]
        movq	qword[x],xmm0
	prtm "m2"

	mov	rbx,[j]
	inc	rbx
	mov	[j],rbx
	cmp 	rbx,[ncol2]	; j<ncol2
	jne	loopj3

	mov	rax,[i]
	inc	rax
	mov	[i],rax
	cmp	rax,[nrow2]	; i<nrow2
	jne	loopi3


	mov	rdi,fmtbl       ; print blank line
	mov	rax,0		; no float to print
        call    printf		; Call C function

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