// SPDX-License-Identifier: GPL-2.0-or-later

/***************************************************************************
 *   Copyright (C) 2013-2014,2019-2020 Synopsys, Inc.                      *
 *   Frank Dols <frank.dols@synopsys.com>                                  *
 *   Mischa Jonker <mischa.jonker@synopsys.com>                            *
 *   Anton Kolesov <anton.kolesov@synopsys.com>                            *
 *   Evgeniy Didin <didin@synopsys.com>                                    *
 ***************************************************************************/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#include "arc.h"

/* ----- Supporting functions ---------------------------------------------- */
static bool arc_mem_is_slow_memory(struct arc_common *arc, uint32_t addr,
	uint32_t size, uint32_t count)
{
	uint32_t addr_end = addr + size * count;
	/* `_end` field can overflow - it points to the first byte after the end,
	 * therefore if DCCM is right at the end of memory address space, then
	 * dccm_end will be 0. */
	assert(addr_end >= addr || addr_end == 0);

	return !((addr >= arc->dccm_start && addr_end <= arc->dccm_end) ||
		(addr >= arc->iccm0_start && addr_end <= arc->iccm0_end) ||
		(addr >= arc->iccm1_start && addr_end <= arc->iccm1_end));
}

/* Write word at word-aligned address */
static int arc_mem_write_block32(struct target *target, uint32_t addr,
	uint32_t count, void *buf)
{
	struct arc_common *arc = target_to_arc(target);

	LOG_DEBUG("Write 4-byte memory block: addr=0x%08" PRIx32 ", count=%" PRIu32,
			addr, count);

	/* Check arguments */
	assert(!(addr & 3));

	/* We need to flush the cache since it might contain dirty
	 * lines, so the cache invalidation may cause data inconsistency. */
	CHECK_RETVAL(arc_cache_flush(target));


	/* No need to flush cache, because we don't read values from memory. */
	CHECK_RETVAL(arc_jtag_write_memory(&arc->jtag_info, addr, count,
				(uint32_t *)buf));

	/* Invalidate caches. */
	CHECK_RETVAL(arc_cache_invalidate(target));

	return ERROR_OK;
}

/* Write half-word at half-word-aligned address */
static int arc_mem_write_block16(struct target *target, uint32_t addr,
	uint32_t count, void *buf)
{
	struct arc_common *arc = target_to_arc(target);
	uint32_t i;
	uint32_t buffer_he;
	uint8_t buffer_te[sizeof(uint32_t)];
	uint8_t halfword_te[sizeof(uint16_t)];

	LOG_DEBUG("Write 2-byte memory block: addr=0x%08" PRIx32 ", count=%" PRIu32,
			addr, count);

	/* Check arguments */
	assert(!(addr & 1));

	/* We will read data from memory, so we need to flush the cache. */
	CHECK_RETVAL(arc_cache_flush(target));

	/* non-word writes are less common than 4-byte writes, so I suppose we can
	 * allow ourselves to write this in a cycle, instead of calling arc_jtag
	 * with count > 1. */
	for (i = 0; i < count; i++) {
		/* We can read only word at word-aligned address. Also *jtag_read_memory
		 * functions return data in host endianness, so host endianness !=
		 * target endianness we have to convert data back to target endianness,
		 * or bytes will be at the wrong places.So:
		 *   1) read word
		 *   2) convert to target endianness
		 *   3) make changes
		 *   4) convert back to host endianness
		 *   5) write word back to target.
		 */
		bool is_slow_memory = arc_mem_is_slow_memory(arc,
			(addr + i * sizeof(uint16_t)) & ~3u, 4, 1);
		CHECK_RETVAL(arc_jtag_read_memory(&arc->jtag_info,
				(addr + i * sizeof(uint16_t)) & ~3u, 1, &buffer_he,
				is_slow_memory));
		target_buffer_set_u32(target, buffer_te, buffer_he);

		/* buf is in host endianness, convert to target */
		target_buffer_set_u16(target, halfword_te, ((uint16_t *)buf)[i]);

		memcpy(buffer_te  + ((addr + i * sizeof(uint16_t)) & 3u),
			halfword_te, sizeof(uint16_t));

		buffer_he = target_buffer_get_u32(target, buffer_te);

		CHECK_RETVAL(arc_jtag_write_memory(&arc->jtag_info,
			(addr + i * sizeof(uint16_t)) & ~3u, 1, &buffer_he));
	}

	/* Invalidate caches. */
	CHECK_RETVAL(arc_cache_invalidate(target));

	return ERROR_OK;
}

/* Write byte at address */
static int arc_mem_write_block8(struct target *target, uint32_t addr,
	uint32_t count, void *buf)
{
	struct arc_common *arc = target_to_arc(target);
	uint32_t i;
	uint32_t buffer_he;
	uint8_t buffer_te[sizeof(uint32_t)];


	LOG_DEBUG("Write 1-byte memory block: addr=0x%08" PRIx32 ", count=%" PRIu32,
			addr, count);

	/* We will read data from memory, so we need to flush the cache. */
	CHECK_RETVAL(arc_cache_flush(target));

	/* non-word writes are less common than 4-byte writes, so I suppose we can
	 * allow ourselves to write this in a cycle, instead of calling arc_jtag
	 * with count > 1. */
	for (i = 0; i < count; i++) {
		/* See comment in arc_mem_write_block16 for details. Since it is a byte
		 * there is not need to convert write buffer to target endianness, but
		 * we still have to convert read buffer. */
		CHECK_RETVAL(arc_jtag_read_memory(&arc->jtag_info, (addr + i) & ~3, 1, &buffer_he,
			    arc_mem_is_slow_memory(arc, (addr + i) & ~3, 4, 1)));
		target_buffer_set_u32(target, buffer_te, buffer_he);
		memcpy(buffer_te  + ((addr + i) & 3), (uint8_t *)buf + i, 1);
		buffer_he = target_buffer_get_u32(target, buffer_te);
		CHECK_RETVAL(arc_jtag_write_memory(&arc->jtag_info, (addr + i) & ~3, 1, &buffer_he));
	}

	/* Invalidate caches. */
	CHECK_RETVAL(arc_cache_invalidate(target));

	return ERROR_OK;
}

/* ----- Exported functions ------------------------------------------------ */
int arc_mem_write(struct target *target, target_addr_t address, uint32_t size,
	uint32_t count, const uint8_t *buffer)
{
	int retval = ERROR_OK;
	void *tunnel = NULL;

	LOG_DEBUG("address: 0x%08" TARGET_PRIxADDR ", size: %" PRIu32 ", count: %" PRIu32,
		address, size, count);

	if (target->state != TARGET_HALTED) {
		LOG_TARGET_ERROR(target, "not halted");
		return ERROR_TARGET_NOT_HALTED;
	}

	/* sanitize arguments */
	if (((size != 4) && (size != 2) && (size != 1)) || !(count) || !(buffer))
		return ERROR_COMMAND_SYNTAX_ERROR;

	if (((size == 4) && (address & 0x3u)) || ((size == 2) && (address & 0x1u)))
		return ERROR_TARGET_UNALIGNED_ACCESS;

	/* correct endianness if we have word or hword access */
	if (size > 1) {
		/*
		 * arc_..._write_mem with size 4/2 requires uint32_t/uint16_t
		 * in host endianness, but byte array represents target endianness.
		 */
		tunnel = calloc(1, count * size * sizeof(uint8_t));

		if (!tunnel) {
			LOG_ERROR("Unable to allocate memory");
			return ERROR_FAIL;
		}

		switch (size) {
		case 4:
			target_buffer_get_u32_array(target, buffer, count,
				(uint32_t *)tunnel);
			break;
		case 2:
			target_buffer_get_u16_array(target, buffer, count,
				(uint16_t *)tunnel);
			break;
		}
		buffer = tunnel;
	}

	if (size == 4) {
		retval = arc_mem_write_block32(target, address, count, (void *)buffer);
	} else if (size == 2) {
		/* We convert buffer from host endianness to target. But then in
		 * write_block16, we do the reverse. Is there a way to avoid this without
		 * breaking other cases? */
		retval = arc_mem_write_block16(target, address, count, (void *)buffer);
	} else {
		retval = arc_mem_write_block8(target, address, count, (void *)buffer);
	}

	free(tunnel);

	return retval;
}

static int arc_mem_read_block(struct target *target, target_addr_t addr,
	uint32_t size, uint32_t count, void *buf)
{
	struct arc_common *arc = target_to_arc(target);

	LOG_DEBUG("Read memory: addr=0x%08" TARGET_PRIxADDR ", size=%" PRIu32
			", count=%" PRIu32, addr, size, count);
	assert(!(addr & 3));
	assert(size == 4);

	/* Flush cache before memory access */
	CHECK_RETVAL(arc_cache_flush(target));

	CHECK_RETVAL(arc_jtag_read_memory(&arc->jtag_info, addr, count, buf,
		    arc_mem_is_slow_memory(arc, addr, size, count)));

	return ERROR_OK;
}

int arc_mem_read(struct target *target, target_addr_t address, uint32_t size,
	uint32_t count, uint8_t *buffer)
{
	int retval = ERROR_OK;
	void *tunnel_he;
	uint8_t *tunnel_te;
	uint32_t words_to_read, bytes_to_read;


	LOG_DEBUG("Read memory: addr=0x%08" TARGET_PRIxADDR ", size=%" PRIu32
			", count=%" PRIu32, address, size, count);

	if (target->state != TARGET_HALTED) {
		LOG_WARNING("target not halted");
		return ERROR_TARGET_NOT_HALTED;
	}

	/* Sanitize arguments */
	if (((size != 4) && (size != 2) && (size != 1)) || !(count) || !(buffer))
		return ERROR_COMMAND_SYNTAX_ERROR;

	if (((size == 4) && (address & 0x3u)) || ((size == 2) && (address & 0x1u)))
	    return ERROR_TARGET_UNALIGNED_ACCESS;

	/* Reads are word-aligned, so padding might be required if count > 1.
	 * NB: +3 is a padding for the last word (in case it's not aligned;
	 * addr&3 is a padding for the first word (since address can be
	 * unaligned as well).  */
	bytes_to_read = (count * size + 3 + (address & 3u)) & ~3u;
	words_to_read = bytes_to_read >> 2;
	tunnel_he = calloc(1, bytes_to_read);
	tunnel_te = calloc(1, bytes_to_read);

	if (!tunnel_he || !tunnel_te) {
		LOG_ERROR("Unable to allocate memory");
		free(tunnel_he);
		free(tunnel_te);
		return ERROR_FAIL;
	}

	/* We can read only word-aligned words. */
	retval = arc_mem_read_block(target, address & ~3u, sizeof(uint32_t),
		words_to_read, tunnel_he);

	/* arc_..._read_mem with size 4/2 returns uint32_t/uint16_t in host */
	/* endianness, but byte array should represent target endianness      */

	if (retval == ERROR_OK) {
		switch (size) {
		case 4:
			target_buffer_set_u32_array(target, buffer, count,
				tunnel_he);
			break;
		case 2:
			target_buffer_set_u32_array(target, tunnel_te,
				words_to_read, tunnel_he);
			/* Will that work properly with count > 1 and big endian? */
			memcpy(buffer, tunnel_te + (address & 3u),
				count * sizeof(uint16_t));
			break;
		case 1:
			target_buffer_set_u32_array(target, tunnel_te,
				words_to_read, tunnel_he);
			/* Will that work properly with count > 1 and big endian? */
			memcpy(buffer, tunnel_te + (address & 3u), count);
			break;
		}
	}

	free(tunnel_he);
	free(tunnel_te);

	return retval;
}