.global sha1_chunk

.text
	counter: .asciz "counter: %ld "
	w1:		.asciz "w[i-3]: %ld "
	w2:		.asciz "w[i-8]: %ld "
	w3:		.asciz "w[i-14]: %ld "
	w4:		.asciz "w[i-16]: %ld "
	xor:	.asciz "XOR: %ld "
	lor:	.asciz "ROTATE: %ld\n"
	offset: .asciz "OFFSET: %ld "
	address: .asciz "ADDRESS: %ld\n"
	test:	.asciz "TEST: %ld\n"

# First parameter (rdi): address of h0
# Second parameter (rsi): address of the first 32-bit word of an array of 80 32-bit words.

sha1_chunk:
	# Setting up subroutine
	pushq	%rbp				# Push the base pointer on the stack
	movq	%rsp, %rbp			# Put the stack pointer into rbp

	pushq	%rdi # -8
	pushq	%rsi # -16

	# Extend the 16 32-bit words into 80 32-bit words
	# for i from 16 to 79
    #    w[i] = (w[i-3] xor w[i-8] xor w[i-14] xor w[i-16]) leftrotate 1

	# Create memory space for loop counter
	pushq	$16

	loopextend:
		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movq	-24(%rbp), %rsi
		movq	$counter, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		# w[i-3]
		movq	-24(%rbp), %r9
		subq	$3, %r9
		movq	$4, %rax
		imulq	%r9, %rax
		addq	%rsi, %rax
		movl	(%rax), %r9d	# address of w[i-3] is now in %r9d

		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movl	%r9d, %esi
		movq	$w1, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		# w[i-8]
		movq	-24(%rbp), %r10
		subq	$8, %r10
		movq	$4, %rax
		imulq	%r10, %rax
		addq	%rsi, %rax
		movl	(%rax), %r10d	# address of w[i-8] is now in %r10d

		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movl	%r10d, %esi
		movq	$w2, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		# w[i-14]
		movq	-24(%rbp), %r11
		subq	$14, %r11
		movq	$4, %rax
		imulq	%r11, %rax
		addq	%rsi, %rax
		movl	(%rax), %r11d	# address of w[i-14] is now in %r11d

		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movl	%r11d, %esi
		movq	$w3, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		# w[i-16]
		movq	-24(%rbp), %r12
		subq	$16, %r12
		movq	$4, %rax
		imulq	%r12, %rax
		addq	%rsi, %rax
		movl	(%rax), %r12d	# address of w[i-14] is now in %r12d

		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movl	%r12d, %esi
		movq	$w4, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		# xor's
		xorl	%r9d, %r10d
		xorl	%r10d, %r11d
		xorl	%r11d, %r12d

		# (w[i-3] xor w[i-8] xor w[i-14] xor w[i-16]) leftrotate 1
		rol		%r12d

		# Assign value
		movq	$4, %rax
		imulq	-24(%rbp), %rax
		addq	%rsi, %rax
		movl	%r12d, (%rax)

		pushq	%rdi
		pushq	%rsi
		movq	$0, %rax
		movl	%r12d, %esi
		movq	$lor, %rdi
		call	printf
		popq	%rsi
		popq	%rdi

		incq	-24(%rbp)
		cmp		$79, -24(%rbp)
		jle		loopextend

	# Print loop counter

	movq	$0, %rax
	movl	64(%rsi), %esi
	movq	$test, %rdi
	call	printf

	# Cleaning up subroutine
	movq	%rbp, %rsp
	popq	%rbp

	ret