archival/libunarchive/decompress_unlzma.c

/* vi: set sw=4 ts=4: */
/*
 * Small lzma deflate implementation.
 * Copyright (C) 2006  Aurelien Jacobs <aurel@gnuage.org>
 *
 * Based on LzmaDecode.c from the LZMA SDK 4.22 (http://www.7-zip.org/)
 * Copyright (C) 1999-2005  Igor Pavlov
 *
 * Licensed under GPLv2 or later, see file LICENSE in this tarball for details.
 */

#include "libbb.h"
#include "unarchive.h"

#ifdef CONFIG_FEATURE_LZMA_FAST
#  define speed_inline ATTRIBUTE_ALWAYS_INLINE
#else
#  define speed_inline
#endif


typedef struct {
	int fd;
	uint8_t *ptr;

/* Was keeping rc on stack in unlzma and separately allocating buffer,
 * but with "buffer 'attached to' allocated rc" code is smaller: */
	/* uint8_t *buffer; */
#define RC_BUFFER ((uint8_t*)(rc+1))

	uint8_t *buffer_end;

/* Had provisions for variable buffer, but we don't need it here */
	/* int buffer_size; */
#define RC_BUFFER_SIZE 0x10000

	uint32_t code;
	uint32_t range;
	uint32_t bound;
} rc_t;

#define RC_TOP_BITS 24
#define RC_MOVE_BITS 5
#define RC_MODEL_TOTAL_BITS 11


/* Called twice: once at startup and once in rc_normalize() */
static void rc_read(rc_t * rc)
{
	int buffer_size = safe_read(rc->fd, RC_BUFFER, RC_BUFFER_SIZE);
	if (buffer_size <= 0)
		bb_error_msg_and_die("unexpected EOF");
	rc->ptr = RC_BUFFER;
	rc->buffer_end = RC_BUFFER + buffer_size;
}

/* Called once */
static rc_t* rc_init(int fd) /*, int buffer_size) */
{
	int i;
	rc_t* rc;

	rc = xmalloc(sizeof(rc_t) + RC_BUFFER_SIZE);

	rc->fd = fd;
	/* rc->buffer_size = buffer_size; */
	rc->buffer_end = RC_BUFFER + RC_BUFFER_SIZE;
	rc->ptr = rc->buffer_end;

	rc->code = 0;
	rc->range = 0xFFFFFFFF;
	for (i = 0; i < 5; i++) {
		if (rc->ptr >= rc->buffer_end)
			rc_read(rc);
		rc->code = (rc->code << 8) | *rc->ptr++;
	}
	return rc;
}

/* Called once  */
static ATTRIBUTE_ALWAYS_INLINE void rc_free(rc_t * rc)
{
	if (ENABLE_FEATURE_CLEAN_UP)
		free(rc);
}

/* Called twice, but one callsite is in speed_inline'd rc_is_bit_0_helper() */
static void rc_do_normalize(rc_t * rc)
{
	if (rc->ptr >= rc->buffer_end)
		rc_read(rc);
	rc->range <<= 8;
	rc->code = (rc->code << 8) | *rc->ptr++;
}
static ATTRIBUTE_ALWAYS_INLINE void rc_normalize(rc_t * rc)
{
	if (rc->range < (1 << RC_TOP_BITS)) {
		rc_do_normalize(rc);
	}
}

/* Called 9 times */
/* Why rc_is_bit_0_helper exists?
 * Because we want to always expose (rc->code < rc->bound) to optimizer
 */
static speed_inline uint32_t rc_is_bit_0_helper(rc_t * rc, uint16_t * p)
{
	rc_normalize(rc);
	rc->bound = *p * (rc->range >> RC_MODEL_TOTAL_BITS);
	return rc->bound;
}
static ATTRIBUTE_ALWAYS_INLINE int rc_is_bit_0(rc_t * rc, uint16_t * p)
{
	uint32_t t = rc_is_bit_0_helper(rc, p);
	return rc->code < t;
}

/* Called ~10 times, but very small, thus inlined */
static speed_inline void rc_update_bit_0(rc_t * rc, uint16_t * p)
{
	rc->range = rc->bound;
	*p += ((1 << RC_MODEL_TOTAL_BITS) - *p) >> RC_MOVE_BITS;
}
static speed_inline void rc_update_bit_1(rc_t * rc, uint16_t * p)
{
	rc->range -= rc->bound;
	rc->code -= rc->bound;
	*p -= *p >> RC_MOVE_BITS;
}

/* Called 4 times in unlzma loop */
static int rc_get_bit(rc_t * rc, uint16_t * p, int *symbol)
{
	if (rc_is_bit_0(rc, p)) {
		rc_update_bit_0(rc, p);
		*symbol *= 2;
		return 0;
	} else {
		rc_update_bit_1(rc, p);
		*symbol = *symbol * 2 + 1;
		return 1;
	}
}

/* Called once */
static ATTRIBUTE_ALWAYS_INLINE int rc_direct_bit(rc_t * rc)
{
	rc_normalize(rc);
	rc->range >>= 1;
	if (rc->code >= rc->range) {
		rc->code -= rc->range;
		return 1;
	}
	return 0;
}

/* Called twice */
static speed_inline void
rc_bit_tree_decode(rc_t * rc, uint16_t * p, int num_levels, int *symbol)
{
	int i = num_levels;

	*symbol = 1;
	while (i--)
		rc_get_bit(rc, p + *symbol, symbol);
	*symbol -= 1 << num_levels;
}


typedef struct {
	uint8_t pos;
	uint32_t dict_size;
	uint64_t dst_size;
} __attribute__ ((packed)) lzma_header_t;


/* #defines will force compiler to compute/optimize each one with each usage.
 * Have heart and use enum instead. */
enum {
	LZMA_BASE_SIZE = 1846,
	LZMA_LIT_SIZE  = 768,

	LZMA_NUM_POS_BITS_MAX = 4,

	LZMA_LEN_NUM_LOW_BITS  = 3,
	LZMA_LEN_NUM_MID_BITS  = 3,
	LZMA_LEN_NUM_HIGH_BITS = 8,

	LZMA_LEN_CHOICE     = 0,
	LZMA_LEN_CHOICE_2   = (LZMA_LEN_CHOICE + 1),
	LZMA_LEN_LOW        = (LZMA_LEN_CHOICE_2 + 1),
	LZMA_LEN_MID        = (LZMA_LEN_LOW \
	                      + (1 << (LZMA_NUM_POS_BITS_MAX + LZMA_LEN_NUM_LOW_BITS))),
	LZMA_LEN_HIGH       = (LZMA_LEN_MID \
	                      + (1 << (LZMA_NUM_POS_BITS_MAX + LZMA_LEN_NUM_MID_BITS))),
	LZMA_NUM_LEN_PROBS  = (LZMA_LEN_HIGH + (1 << LZMA_LEN_NUM_HIGH_BITS)),

	LZMA_NUM_STATES     = 12,
	LZMA_NUM_LIT_STATES = 7,

	LZMA_START_POS_MODEL_INDEX = 4,
	LZMA_END_POS_MODEL_INDEX   = 14,
	LZMA_NUM_FULL_DISTANCES    = (1 << (LZMA_END_POS_MODEL_INDEX >> 1)),

	LZMA_NUM_POS_SLOT_BITS = 6,
	LZMA_NUM_LEN_TO_POS_STATES = 4,

	LZMA_NUM_ALIGN_BITS = 4,

	LZMA_MATCH_MIN_LEN  = 2,

	LZMA_IS_MATCH       = 0,
	LZMA_IS_REP         = (LZMA_IS_MATCH + (LZMA_NUM_STATES << LZMA_NUM_POS_BITS_MAX)),
	LZMA_IS_REP_G0      = (LZMA_IS_REP + LZMA_NUM_STATES),
	LZMA_IS_REP_G1      = (LZMA_IS_REP_G0 + LZMA_NUM_STATES),
	LZMA_IS_REP_G2      = (LZMA_IS_REP_G1 + LZMA_NUM_STATES),
	LZMA_IS_REP_0_LONG  = (LZMA_IS_REP_G2 + LZMA_NUM_STATES),
	LZMA_POS_SLOT       = (LZMA_IS_REP_0_LONG \
	                      + (LZMA_NUM_STATES << LZMA_NUM_POS_BITS_MAX)),
	LZMA_SPEC_POS       = (LZMA_POS_SLOT \
	                      + (LZMA_NUM_LEN_TO_POS_STATES << LZMA_NUM_POS_SLOT_BITS)),
	LZMA_ALIGN          = (LZMA_SPEC_POS \
	                      + LZMA_NUM_FULL_DISTANCES - LZMA_END_POS_MODEL_INDEX),
	LZMA_LEN_CODER      = (LZMA_ALIGN + (1 << LZMA_NUM_ALIGN_BITS)),
	LZMA_REP_LEN_CODER  = (LZMA_LEN_CODER + LZMA_NUM_LEN_PROBS),
	LZMA_LITERAL        = (LZMA_REP_LEN_CODER + LZMA_NUM_LEN_PROBS),
};


USE_DESKTOP(long long) int
unlzma(int src_fd, int dst_fd)
{
	USE_DESKTOP(long long total_written = 0;)
	lzma_header_t header;
	int lc, pb, lp;
	uint32_t pos_state_mask;
	uint32_t literal_pos_mask;
	uint32_t pos;
	uint16_t *p;
	uint16_t *prob;
	uint16_t *prob_lit;
	int num_bits;
	int num_probs;
	rc_t *rc;
	int i, mi;
	uint8_t *buffer;
	uint8_t previous_byte = 0;
	size_t buffer_pos = 0, global_pos = 0;
	int len = 0;
	int state = 0;
	uint32_t rep0 = 1, rep1 = 1, rep2 = 1, rep3 = 1;

	xread(src_fd, &header, sizeof(header));

	if (header.pos >= (9 * 5 * 5))
		bb_error_msg_and_die("bad header");
	mi = header.pos / 9;
	lc = header.pos % 9;
	pb = mi / 5;
	lp = mi % 5;
	pos_state_mask = (1 << pb) - 1;
	literal_pos_mask = (1 << lp) - 1;

	header.dict_size = SWAP_LE32(header.dict_size);
	header.dst_size = SWAP_LE64(header.dst_size);

	if (header.dict_size == 0)
		header.dict_size = 1;

	buffer = xmalloc(MIN(header.dst_size, header.dict_size));

	num_probs = LZMA_BASE_SIZE + (LZMA_LIT_SIZE << (lc + lp));
	p = xmalloc(num_probs * sizeof(*p));
	num_probs = LZMA_LITERAL + (LZMA_LIT_SIZE << (lc + lp));
	for (i = 0; i < num_probs; i++)
		p[i] = (1 << RC_MODEL_TOTAL_BITS) >> 1;

	rc = rc_init(src_fd); /*, RC_BUFFER_SIZE); */

	while (global_pos + buffer_pos < header.dst_size) {
		int pos_state = (buffer_pos + global_pos) & pos_state_mask;

		prob =
			p + LZMA_IS_MATCH + (state << LZMA_NUM_POS_BITS_MAX) + pos_state;
		if (rc_is_bit_0(rc, prob)) {
			mi = 1;
			rc_update_bit_0(rc, prob);
			prob = (p + LZMA_LITERAL + (LZMA_LIT_SIZE
					* ((((buffer_pos + global_pos) & literal_pos_mask) << lc)
					+ (previous_byte >> (8 - lc)))));

			if (state >= LZMA_NUM_LIT_STATES) {
				int match_byte;

				pos = buffer_pos - rep0;
				while (pos >= header.dict_size)
					pos += header.dict_size;
				match_byte = buffer[pos];
				do {
					int bit;

					match_byte <<= 1;
					bit = match_byte & 0x100;
					prob_lit = prob + 0x100 + bit + mi;
					if (rc_get_bit(rc, prob_lit, &mi)) {
						if (!bit)
							break;
					} else {
						if (bit)
							break;
					}
				} while (mi < 0x100);
			}
			while (mi < 0x100) {
				prob_lit = prob + mi;
				rc_get_bit(rc, prob_lit, &mi);
			}
			previous_byte = (uint8_t) mi;

			buffer[buffer_pos++] = previous_byte;
			if (buffer_pos == header.dict_size) {
				buffer_pos = 0;
				global_pos += header.dict_size;
				if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
					goto bad;
				USE_DESKTOP(total_written += header.dict_size;)
			}
			if (state < 4)
				state = 0;
			else if (state < 10)
				state -= 3;
			else
				state -= 6;
		} else {
			int offset;
			uint16_t *prob_len;

			rc_update_bit_1(rc, prob);
			prob = p + LZMA_IS_REP + state;
			if (rc_is_bit_0(rc, prob)) {
				rc_update_bit_0(rc, prob);
				rep3 = rep2;
				rep2 = rep1;
				rep1 = rep0;
				state = state < LZMA_NUM_LIT_STATES ? 0 : 3;
				prob = p + LZMA_LEN_CODER;
			} else {
				rc_update_bit_1(rc, prob);
				prob = p + LZMA_IS_REP_G0 + state;
				if (rc_is_bit_0(rc, prob)) {
					rc_update_bit_0(rc, prob);
					prob = (p + LZMA_IS_REP_0_LONG
							+ (state << LZMA_NUM_POS_BITS_MAX) + pos_state);
					if (rc_is_bit_0(rc, prob)) {
						rc_update_bit_0(rc, prob);

						state = state < LZMA_NUM_LIT_STATES ? 9 : 11;
						pos = buffer_pos - rep0;
						while (pos >= header.dict_size)
							pos += header.dict_size;
						previous_byte = buffer[pos];
						buffer[buffer_pos++] = previous_byte;
						if (buffer_pos == header.dict_size) {
							buffer_pos = 0;
							global_pos += header.dict_size;
							if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
								goto bad;
							USE_DESKTOP(total_written += header.dict_size;)
						}
						continue;
					} else {
						rc_update_bit_1(rc, prob);
					}
				} else {
					uint32_t distance;

					rc_update_bit_1(rc, prob);
					prob = p + LZMA_IS_REP_G1 + state;
					if (rc_is_bit_0(rc, prob)) {
						rc_update_bit_0(rc, prob);
						distance = rep1;
					} else {
						rc_update_bit_1(rc, prob);
						prob = p + LZMA_IS_REP_G2 + state;
						if (rc_is_bit_0(rc, prob)) {
							rc_update_bit_0(rc, prob);
							distance = rep2;
						} else {
							rc_update_bit_1(rc, prob);
							distance = rep3;
							rep3 = rep2;
						}
						rep2 = rep1;
					}
					rep1 = rep0;
					rep0 = distance;
				}
				state = state < LZMA_NUM_LIT_STATES ? 8 : 11;
				prob = p + LZMA_REP_LEN_CODER;
			}

			prob_len = prob + LZMA_LEN_CHOICE;
			if (rc_is_bit_0(rc, prob_len)) {
				rc_update_bit_0(rc, prob_len);
				prob_len = (prob + LZMA_LEN_LOW
							+ (pos_state << LZMA_LEN_NUM_LOW_BITS));
				offset = 0;
				num_bits = LZMA_LEN_NUM_LOW_BITS;
			} else {
				rc_update_bit_1(rc, prob_len);
				prob_len = prob + LZMA_LEN_CHOICE_2;
				if (rc_is_bit_0(rc, prob_len)) {
					rc_update_bit_0(rc, prob_len);
					prob_len = (prob + LZMA_LEN_MID
								+ (pos_state << LZMA_LEN_NUM_MID_BITS));
					offset = 1 << LZMA_LEN_NUM_LOW_BITS;
					num_bits = LZMA_LEN_NUM_MID_BITS;
				} else {
					rc_update_bit_1(rc, prob_len);
					prob_len = prob + LZMA_LEN_HIGH;
					offset = ((1 << LZMA_LEN_NUM_LOW_BITS)
							  + (1 << LZMA_LEN_NUM_MID_BITS));
					num_bits = LZMA_LEN_NUM_HIGH_BITS;
				}
			}
			rc_bit_tree_decode(rc, prob_len, num_bits, &len);
			len += offset;

			if (state < 4) {
				int pos_slot;

				state += LZMA_NUM_LIT_STATES;
				prob =
					p + LZMA_POS_SLOT +
					((len <
					  LZMA_NUM_LEN_TO_POS_STATES ? len :
					  LZMA_NUM_LEN_TO_POS_STATES - 1)
					 << LZMA_NUM_POS_SLOT_BITS);
				rc_bit_tree_decode(rc, prob, LZMA_NUM_POS_SLOT_BITS,
								   &pos_slot);
				if (pos_slot >= LZMA_START_POS_MODEL_INDEX) {
					num_bits = (pos_slot >> 1) - 1;
					rep0 = 2 | (pos_slot & 1);
					if (pos_slot < LZMA_END_POS_MODEL_INDEX) {
						rep0 <<= num_bits;
						prob = p + LZMA_SPEC_POS + rep0 - pos_slot - 1;
					} else {
						num_bits -= LZMA_NUM_ALIGN_BITS;
						while (num_bits--)
							rep0 = (rep0 << 1) | rc_direct_bit(rc);
						prob = p + LZMA_ALIGN;
						rep0 <<= LZMA_NUM_ALIGN_BITS;
						num_bits = LZMA_NUM_ALIGN_BITS;
					}
					i = 1;
					mi = 1;
					while (num_bits--) {
						if (rc_get_bit(rc, prob + mi, &mi))
							rep0 |= i;
						i <<= 1;
					}
				} else
					rep0 = pos_slot;
				if (++rep0 == 0)
					break;
			}

			len += LZMA_MATCH_MIN_LEN;

			do {
				pos = buffer_pos - rep0;
				while (pos >= header.dict_size)
					pos += header.dict_size;
				previous_byte = buffer[pos];
				buffer[buffer_pos++] = previous_byte;
				if (buffer_pos == header.dict_size) {
					buffer_pos = 0;
					global_pos += header.dict_size;
					if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
						goto bad;
					USE_DESKTOP(total_written += header.dict_size;)
				}
				len--;
			} while (len != 0 && buffer_pos < header.dst_size);
		}
	}


	if (full_write(dst_fd, buffer, buffer_pos) != buffer_pos) {
 bad:
		rc_free(rc);
		return -1;
	}
	rc_free(rc);
	USE_DESKTOP(total_written += buffer_pos;)
	return USE_DESKTOP(total_written) + 0;
}
1	/* vi: set sw=4 ts=4: */
2	/*
3	* Small lzma deflate implementation.
4	* Copyright (C) 2006 Aurelien Jacobs <aurel@gnuage.org>
5	*
6	* Based on LzmaDecode.c from the LZMA SDK 4.22 (http://www.7-zip.org/)
7	* Copyright (C) 1999-2005 Igor Pavlov
8	*
9	* Licensed under GPLv2 or later, see file LICENSE in this tarball for details.
10	*/
11
12	#include "libbb.h"
13	#include "unarchive.h"
14
15	#ifdef CONFIG_FEATURE_LZMA_FAST
16	# define speed_inline ATTRIBUTE_ALWAYS_INLINE
17	#else
18	# define speed_inline
19	#endif
20
21
22	typedef struct {
23	int fd;
24	uint8_t *ptr;
25
26	/* Was keeping rc on stack in unlzma and separately allocating buffer,
27	* but with "buffer 'attached to' allocated rc" code is smaller: */
28	/* uint8_t buffer; /
29	#define RC_BUFFER ((uint8_t*)(rc+1))
30
31	uint8_t *buffer_end;
32
33	/* Had provisions for variable buffer, but we don't need it here */
34	/* int buffer_size; */
35	#define RC_BUFFER_SIZE 0x10000
36
37	uint32_t code;
38	uint32_t range;
39	uint32_t bound;
40	} rc_t;
41
42	#define RC_TOP_BITS 24
43	#define RC_MOVE_BITS 5
44	#define RC_MODEL_TOTAL_BITS 11
45
46
47	/* Called twice: once at startup and once in rc_normalize() */
48	static void rc_read(rc_t * rc)
49	{
50	int buffer_size = safe_read(rc->fd, RC_BUFFER, RC_BUFFER_SIZE);
51	if (buffer_size <= 0)
52	bb_error_msg_and_die("unexpected EOF");
53	rc->ptr = RC_BUFFER;
54	rc->buffer_end = RC_BUFFER + buffer_size;
55	}
56
57	/* Called once */
58	static rc_t* rc_init(int fd) /, int buffer_size) /
59	{
60	int i;
61	rc_t* rc;
62
63	rc = xmalloc(sizeof(rc_t) + RC_BUFFER_SIZE);
64
65	rc->fd = fd;
66	/* rc->buffer_size = buffer_size; */
67	rc->buffer_end = RC_BUFFER + RC_BUFFER_SIZE;
68	rc->ptr = rc->buffer_end;
69
70	rc->code = 0;
71	rc->range = 0xFFFFFFFF;
72	for (i = 0; i < 5; i++) {
73	if (rc->ptr >= rc->buffer_end)
74	rc_read(rc);
75	rc->code = (rc->code << 8) \| *rc->ptr++;
76	}
77	return rc;
78	}
79
80	/* Called once */
81	static ATTRIBUTE_ALWAYS_INLINE void rc_free(rc_t * rc)
82	{
83	if (ENABLE_FEATURE_CLEAN_UP)
84	free(rc);
85	}
86
87	/* Called twice, but one callsite is in speed_inline'd rc_is_bit_0_helper() */
88	static void rc_do_normalize(rc_t * rc)
89	{
90	if (rc->ptr >= rc->buffer_end)
91	rc_read(rc);
92	rc->range <<= 8;
93	rc->code = (rc->code << 8) \| *rc->ptr++;
94	}
95	static ATTRIBUTE_ALWAYS_INLINE void rc_normalize(rc_t * rc)
96	{
97	if (rc->range < (1 << RC_TOP_BITS)) {
98	rc_do_normalize(rc);
99	}
100	}
101
102	/* Called 9 times */
103	/* Why rc_is_bit_0_helper exists?
104	* Because we want to always expose (rc->code < rc->bound) to optimizer
105	*/
106	static speed_inline uint32_t rc_is_bit_0_helper(rc_t * rc, uint16_t * p)
107	{
108	rc_normalize(rc);
109	rc->bound = p (rc->range >> RC_MODEL_TOTAL_BITS);
110	return rc->bound;
111	}
112	static ATTRIBUTE_ALWAYS_INLINE int rc_is_bit_0(rc_t * rc, uint16_t * p)
113	{
114	uint32_t t = rc_is_bit_0_helper(rc, p);
115	return rc->code < t;
116	}
117
118	/* Called ~10 times, but very small, thus inlined */
119	static speed_inline void rc_update_bit_0(rc_t * rc, uint16_t * p)
120	{
121	rc->range = rc->bound;
122	p += ((1 << RC_MODEL_TOTAL_BITS) - p) >> RC_MOVE_BITS;
123	}
124	static speed_inline void rc_update_bit_1(rc_t * rc, uint16_t * p)
125	{
126	rc->range -= rc->bound;
127	rc->code -= rc->bound;
128	p -= p >> RC_MOVE_BITS;
129	}
130
131	/* Called 4 times in unlzma loop */
132	static int rc_get_bit(rc_t * rc, uint16_t * p, int *symbol)
133	{
134	if (rc_is_bit_0(rc, p)) {
135	rc_update_bit_0(rc, p);
136	symbol = 2;
137	return 0;
138	} else {
139	rc_update_bit_1(rc, p);
140	symbol = symbol * 2 + 1;
141	return 1;
142	}
143	}
144
145	/* Called once */
146	static ATTRIBUTE_ALWAYS_INLINE int rc_direct_bit(rc_t * rc)
147	{
148	rc_normalize(rc);
149	rc->range >>= 1;
150	if (rc->code >= rc->range) {
151	rc->code -= rc->range;
152	return 1;
153	}
154	return 0;
155	}
156
157	/* Called twice */
158	static speed_inline void
159	rc_bit_tree_decode(rc_t * rc, uint16_t * p, int num_levels, int *symbol)
160	{
161	int i = num_levels;
162
163	*symbol = 1;
164	while (i--)
165	rc_get_bit(rc, p + *symbol, symbol);
166	*symbol -= 1 << num_levels;
167	}
168
169
170	typedef struct {
171	uint8_t pos;
172	uint32_t dict_size;
173	uint64_t dst_size;
174	} __attribute__ ((packed)) lzma_header_t;
175
176
177	/* #defines will force compiler to compute/optimize each one with each usage.
178	* Have heart and use enum instead. */
179	enum {
180	LZMA_BASE_SIZE = 1846,
181	LZMA_LIT_SIZE = 768,
182
183	LZMA_NUM_POS_BITS_MAX = 4,
184
185	LZMA_LEN_NUM_LOW_BITS = 3,
186	LZMA_LEN_NUM_MID_BITS = 3,
187	LZMA_LEN_NUM_HIGH_BITS = 8,
188
189	LZMA_LEN_CHOICE = 0,
190	LZMA_LEN_CHOICE_2 = (LZMA_LEN_CHOICE + 1),
191	LZMA_LEN_LOW = (LZMA_LEN_CHOICE_2 + 1),
192	LZMA_LEN_MID = (LZMA_LEN_LOW \
193	+ (1 << (LZMA_NUM_POS_BITS_MAX + LZMA_LEN_NUM_LOW_BITS))),
194	LZMA_LEN_HIGH = (LZMA_LEN_MID \
195	+ (1 << (LZMA_NUM_POS_BITS_MAX + LZMA_LEN_NUM_MID_BITS))),
196	LZMA_NUM_LEN_PROBS = (LZMA_LEN_HIGH + (1 << LZMA_LEN_NUM_HIGH_BITS)),
197
198	LZMA_NUM_STATES = 12,
199	LZMA_NUM_LIT_STATES = 7,
200
201	LZMA_START_POS_MODEL_INDEX = 4,
202	LZMA_END_POS_MODEL_INDEX = 14,
203	LZMA_NUM_FULL_DISTANCES = (1 << (LZMA_END_POS_MODEL_INDEX >> 1)),
204
205	LZMA_NUM_POS_SLOT_BITS = 6,
206	LZMA_NUM_LEN_TO_POS_STATES = 4,
207
208	LZMA_NUM_ALIGN_BITS = 4,
209
210	LZMA_MATCH_MIN_LEN = 2,
211
212	LZMA_IS_MATCH = 0,
213	LZMA_IS_REP = (LZMA_IS_MATCH + (LZMA_NUM_STATES << LZMA_NUM_POS_BITS_MAX)),
214	LZMA_IS_REP_G0 = (LZMA_IS_REP + LZMA_NUM_STATES),
215	LZMA_IS_REP_G1 = (LZMA_IS_REP_G0 + LZMA_NUM_STATES),
216	LZMA_IS_REP_G2 = (LZMA_IS_REP_G1 + LZMA_NUM_STATES),
217	LZMA_IS_REP_0_LONG = (LZMA_IS_REP_G2 + LZMA_NUM_STATES),
218	LZMA_POS_SLOT = (LZMA_IS_REP_0_LONG \
219	+ (LZMA_NUM_STATES << LZMA_NUM_POS_BITS_MAX)),
220	LZMA_SPEC_POS = (LZMA_POS_SLOT \
221	+ (LZMA_NUM_LEN_TO_POS_STATES << LZMA_NUM_POS_SLOT_BITS)),
222	LZMA_ALIGN = (LZMA_SPEC_POS \
223	+ LZMA_NUM_FULL_DISTANCES - LZMA_END_POS_MODEL_INDEX),
224	LZMA_LEN_CODER = (LZMA_ALIGN + (1 << LZMA_NUM_ALIGN_BITS)),
225	LZMA_REP_LEN_CODER = (LZMA_LEN_CODER + LZMA_NUM_LEN_PROBS),
226	LZMA_LITERAL = (LZMA_REP_LEN_CODER + LZMA_NUM_LEN_PROBS),
227	};
228
229
230	USE_DESKTOP(long long) int
231	unlzma(int src_fd, int dst_fd)
232	{
233	USE_DESKTOP(long long total_written = 0;)
234	lzma_header_t header;
235	int lc, pb, lp;
236	uint32_t pos_state_mask;
237	uint32_t literal_pos_mask;
238	uint32_t pos;
239	uint16_t *p;
240	uint16_t *prob;
241	uint16_t *prob_lit;
242	int num_bits;
243	int num_probs;
244	rc_t *rc;
245	int i, mi;
246	uint8_t *buffer;
247	uint8_t previous_byte = 0;
248	size_t buffer_pos = 0, global_pos = 0;
249	int len = 0;
250	int state = 0;
251	uint32_t rep0 = 1, rep1 = 1, rep2 = 1, rep3 = 1;
252
253	xread(src_fd, &header, sizeof(header));
254
255	if (header.pos >= (9 * 5 * 5))
256	bb_error_msg_and_die("bad header");
257	mi = header.pos / 9;
258	lc = header.pos % 9;
259	pb = mi / 5;
260	lp = mi % 5;
261	pos_state_mask = (1 << pb) - 1;
262	literal_pos_mask = (1 << lp) - 1;
263
264	header.dict_size = SWAP_LE32(header.dict_size);
265	header.dst_size = SWAP_LE64(header.dst_size);
266
267	if (header.dict_size == 0)
268	header.dict_size = 1;
269
270	buffer = xmalloc(MIN(header.dst_size, header.dict_size));
271
272	num_probs = LZMA_BASE_SIZE + (LZMA_LIT_SIZE << (lc + lp));
273	p = xmalloc(num_probs * sizeof(*p));
274	num_probs = LZMA_LITERAL + (LZMA_LIT_SIZE << (lc + lp));
275	for (i = 0; i < num_probs; i++)
276	p[i] = (1 << RC_MODEL_TOTAL_BITS) >> 1;
277
278	rc = rc_init(src_fd); /, RC_BUFFER_SIZE); /
279
280	while (global_pos + buffer_pos < header.dst_size) {
281	int pos_state = (buffer_pos + global_pos) & pos_state_mask;
282
283	prob =
284	p + LZMA_IS_MATCH + (state << LZMA_NUM_POS_BITS_MAX) + pos_state;
285	if (rc_is_bit_0(rc, prob)) {
286	mi = 1;
287	rc_update_bit_0(rc, prob);
288	prob = (p + LZMA_LITERAL + (LZMA_LIT_SIZE
289	* ((((buffer_pos + global_pos) & literal_pos_mask) << lc)
290	+ (previous_byte >> (8 - lc)))));
291
292	if (state >= LZMA_NUM_LIT_STATES) {
293	int match_byte;
294
295	pos = buffer_pos - rep0;
296	while (pos >= header.dict_size)
297	pos += header.dict_size;
298	match_byte = buffer[pos];
299	do {
300	int bit;
301
302	match_byte <<= 1;
303	bit = match_byte & 0x100;
304	prob_lit = prob + 0x100 + bit + mi;
305	if (rc_get_bit(rc, prob_lit, &mi)) {
306	if (!bit)
307	break;
308	} else {
309	if (bit)
310	break;
311	}
312	} while (mi < 0x100);
313	}
314	while (mi < 0x100) {
315	prob_lit = prob + mi;
316	rc_get_bit(rc, prob_lit, &mi);
317	}
318	previous_byte = (uint8_t) mi;
319
320	buffer[buffer_pos++] = previous_byte;
321	if (buffer_pos == header.dict_size) {
322	buffer_pos = 0;
323	global_pos += header.dict_size;
324	if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
325	goto bad;
326	USE_DESKTOP(total_written += header.dict_size;)
327	}
328	if (state < 4)
329	state = 0;
330	else if (state < 10)
331	state -= 3;
332	else
333	state -= 6;
334	} else {
335	int offset;
336	uint16_t *prob_len;
337
338	rc_update_bit_1(rc, prob);
339	prob = p + LZMA_IS_REP + state;
340	if (rc_is_bit_0(rc, prob)) {
341	rc_update_bit_0(rc, prob);
342	rep3 = rep2;
343	rep2 = rep1;
344	rep1 = rep0;
345	state = state < LZMA_NUM_LIT_STATES ? 0 : 3;
346	prob = p + LZMA_LEN_CODER;
347	} else {
348	rc_update_bit_1(rc, prob);
349	prob = p + LZMA_IS_REP_G0 + state;
350	if (rc_is_bit_0(rc, prob)) {
351	rc_update_bit_0(rc, prob);
352	prob = (p + LZMA_IS_REP_0_LONG
353	+ (state << LZMA_NUM_POS_BITS_MAX) + pos_state);
354	if (rc_is_bit_0(rc, prob)) {
355	rc_update_bit_0(rc, prob);
356
357	state = state < LZMA_NUM_LIT_STATES ? 9 : 11;
358	pos = buffer_pos - rep0;
359	while (pos >= header.dict_size)
360	pos += header.dict_size;
361	previous_byte = buffer[pos];
362	buffer[buffer_pos++] = previous_byte;
363	if (buffer_pos == header.dict_size) {
364	buffer_pos = 0;
365	global_pos += header.dict_size;
366	if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
367	goto bad;
368	USE_DESKTOP(total_written += header.dict_size;)
369	}
370	continue;
371	} else {
372	rc_update_bit_1(rc, prob);
373	}
374	} else {
375	uint32_t distance;
376
377	rc_update_bit_1(rc, prob);
378	prob = p + LZMA_IS_REP_G1 + state;
379	if (rc_is_bit_0(rc, prob)) {
380	rc_update_bit_0(rc, prob);
381	distance = rep1;
382	} else {
383	rc_update_bit_1(rc, prob);
384	prob = p + LZMA_IS_REP_G2 + state;
385	if (rc_is_bit_0(rc, prob)) {
386	rc_update_bit_0(rc, prob);
387	distance = rep2;
388	} else {
389	rc_update_bit_1(rc, prob);
390	distance = rep3;
391	rep3 = rep2;
392	}
393	rep2 = rep1;
394	}
395	rep1 = rep0;
396	rep0 = distance;
397	}
398	state = state < LZMA_NUM_LIT_STATES ? 8 : 11;
399	prob = p + LZMA_REP_LEN_CODER;
400	}
401
402	prob_len = prob + LZMA_LEN_CHOICE;
403	if (rc_is_bit_0(rc, prob_len)) {
404	rc_update_bit_0(rc, prob_len);
405	prob_len = (prob + LZMA_LEN_LOW
406	+ (pos_state << LZMA_LEN_NUM_LOW_BITS));
407	offset = 0;
408	num_bits = LZMA_LEN_NUM_LOW_BITS;
409	} else {
410	rc_update_bit_1(rc, prob_len);
411	prob_len = prob + LZMA_LEN_CHOICE_2;
412	if (rc_is_bit_0(rc, prob_len)) {
413	rc_update_bit_0(rc, prob_len);
414	prob_len = (prob + LZMA_LEN_MID
415	+ (pos_state << LZMA_LEN_NUM_MID_BITS));
416	offset = 1 << LZMA_LEN_NUM_LOW_BITS;
417	num_bits = LZMA_LEN_NUM_MID_BITS;
418	} else {
419	rc_update_bit_1(rc, prob_len);
420	prob_len = prob + LZMA_LEN_HIGH;
421	offset = ((1 << LZMA_LEN_NUM_LOW_BITS)
422	+ (1 << LZMA_LEN_NUM_MID_BITS));
423	num_bits = LZMA_LEN_NUM_HIGH_BITS;
424	}
425	}
426	rc_bit_tree_decode(rc, prob_len, num_bits, &len);
427	len += offset;
428
429	if (state < 4) {
430	int pos_slot;
431
432	state += LZMA_NUM_LIT_STATES;
433	prob =
434	p + LZMA_POS_SLOT +
435	((len <
436	LZMA_NUM_LEN_TO_POS_STATES ? len :
437	LZMA_NUM_LEN_TO_POS_STATES - 1)
438	<< LZMA_NUM_POS_SLOT_BITS);
439	rc_bit_tree_decode(rc, prob, LZMA_NUM_POS_SLOT_BITS,
440	&pos_slot);
441	if (pos_slot >= LZMA_START_POS_MODEL_INDEX) {
442	num_bits = (pos_slot >> 1) - 1;
443	rep0 = 2 \| (pos_slot & 1);
444	if (pos_slot < LZMA_END_POS_MODEL_INDEX) {
445	rep0 <<= num_bits;
446	prob = p + LZMA_SPEC_POS + rep0 - pos_slot - 1;
447	} else {
448	num_bits -= LZMA_NUM_ALIGN_BITS;
449	while (num_bits--)
450	rep0 = (rep0 << 1) \| rc_direct_bit(rc);
451	prob = p + LZMA_ALIGN;
452	rep0 <<= LZMA_NUM_ALIGN_BITS;
453	num_bits = LZMA_NUM_ALIGN_BITS;
454	}
455	i = 1;
456	mi = 1;
457	while (num_bits--) {
458	if (rc_get_bit(rc, prob + mi, &mi))
459	rep0 \|= i;
460	i <<= 1;
461	}
462	} else
463	rep0 = pos_slot;
464	if (++rep0 == 0)
465	break;
466	}
467
468	len += LZMA_MATCH_MIN_LEN;
469
470	do {
471	pos = buffer_pos - rep0;
472	while (pos >= header.dict_size)
473	pos += header.dict_size;
474	previous_byte = buffer[pos];
475	buffer[buffer_pos++] = previous_byte;
476	if (buffer_pos == header.dict_size) {
477	buffer_pos = 0;
478	global_pos += header.dict_size;
479	if (full_write(dst_fd, buffer, header.dict_size) != header.dict_size)
480	goto bad;
481	USE_DESKTOP(total_written += header.dict_size;)
482	}
483	len--;
484	} while (len != 0 && buffer_pos < header.dst_size);
485	}
486	}
487
488
489	if (full_write(dst_fd, buffer, buffer_pos) != buffer_pos) {
490	bad:
491	rc_free(rc);
492	return -1;
493	}
494	rc_free(rc);
495	USE_DESKTOP(total_written += buffer_pos;)
496	return USE_DESKTOP(total_written) + 0;
497	}