/* Copyright (C) 2021,2022 fef <owo@fef.moe>.  All rights reserved. */

#include <asm/common.h>

/* void *memset(void *dest, int c, usize n) */
ENTRY(memset)
	pushl	%ebp
	movl	%esp, %ebp

	cld

	movl	8(%ebp), %edi	/* edi = dest */
	movl	12(%ebp), %eax	/* eax = c */
	movl	16(%ebp), %ecx	/* ecx = n */

	cmpl	$16, %ecx
	jb	out		/* n < 16, not worth the effort */

	testl	$1, %edi
	jz	2f
	stosb
	decl	%ecx

	/* 2-byte aligned */
2:	movl	%eax, %edx
	shll	$8, %eax
	orl	%edx, %eax	/* c |= (c << 8) */

	testl	$2, %edi
	jz	4f
	stosw
	subl	$2, %ecx	/* n -= 2 */

	/* 4-byte aligned, now we can fire stosl */
4:	movl	%eax, %edx
	shll	$16, %eax
	orl	%edx, %eax	/* c |= (c << 16) */
	movl	%ecx, %edx
	andl	$7, %edx	/* edx = n % 8 */
	shrl	$2, %ecx	/* n /= 4 */
	rep
	stosl
	movl	%edx, %ecx

	/* write out remaining bytes (or do the whole memset, if n < 16) */
out:	rep
	stosb
	movl	8(%ebp), %eax	/* return original pointer to dest */
	popl	%ebp
	ret
END(memset)