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

#pragma once

#include <gay/types.h>

#define __HAVE_ARCH_MEMSET
extern void *memset(void *dest, int c, usize n);

#define __HAVE_ARCH_MEMSET16
static inline void *memset16(u16 *dest, u16 val, usize nbyte)
{
	void *dst0 = dest;
	nbyte /= 2;

	__asm__ volatile(
"	rep	\n"
"	stosw	\n"
	: "+c"(nbyte), "+D"(dest)
	: "a"(val)
	: "memory"
	);

	return dst0;
}

#define __HAVE_ARCH_MEMSET32
static inline void *memset32(u32 *dest, u32 val, usize nbyte)
{
	void *dst0 = dest;
	nbyte /= 4;

	__asm__ volatile(
"	rep	\n"
"	stosl	\n"
	: "+c"(nbyte), "+D"(dest)
	: "a"(val)
	: "memory"
	);

	return dst0;
}

#define __HAVE_ARCH_MEMSET64
static inline void *memset64(u64 *dest, u64 val, usize nbyte)
{
	void *dst0 = dest;
	nbyte /= 8;

#ifdef __x86_64__
	__asm__ volatile(
"	rep	\n"
"	stosq	\n"
	: "+c"(nbyte), "+D"(dest)
	: "a"(val)
	: "memory"
	);
#else
	__asm__ volatile(
"1:	stosl		\n"
"	xchgl	%3, %2	\n"
"	stosl		\n"
"	xchgl	%3, %2	\n"
"	loop	1b	\n"
	: "+c"(nbyte), "+D"(dest)
	: "a"((u32)val), "r"((u32)(val >> 32))
	: "memory"
	);
#endif

	return dst0;
}