unlzma: speedup, by Pascal Bellard (pascal.bellard AT ads-lu.com)

LZMA_FAST off: 10% faster and:
function                                             old     new   delta
rc_is_bit_1                                            -      92     +92
rc_do_normalize                                       45      81     +36
rc_update_bit_1                                       25       -     -25
rc_update_bit_0                                       30       -     -30
rc_is_bit_0_helper                                    38       -     -38
rc_get_bit                                            60      17     -43
rc_read                                               48       -     -48
unpack_lzma_stream                                  1768    1517    -251
------------------------------------------------------------------------------
(add/remove: 1/4 grow/shrink: 1/2 up/down: 128/-435)         Total: -307 bytes

LZMA_FAST on: 14% faster and:
unpack_lzma_stream                                  2301    2737    +436
rc_get_bit                                           106       -    -106
This commit is contained in:
Denis Vlasenko 2009-04-25 12:19:35 +00:00
parent 6fb5f01c0b
commit 9ac3dc764a
2 changed files with 60 additions and 98 deletions

View File

@ -283,8 +283,8 @@ config FEATURE_LZMA_FAST
default n
depends on UNLZMA
help
This option reduces decompression time by about 33% at the cost of
a 2K bigger binary.
This option reduces decompression time by about 25% at the cost of
a 1K bigger binary.
config UNZIP
bool "unzip"

View File

@ -8,14 +8,15 @@
*
* Licensed under GPLv2 or later, see file LICENSE in this tarball for details.
*/
#include "libbb.h"
#include "unarchive.h"
#if ENABLE_FEATURE_LZMA_FAST
# define speed_inline ALWAYS_INLINE
# define size_inline
#else
# define speed_inline
# define size_inline ALWAYS_INLINE
#endif
@ -44,8 +45,8 @@ typedef struct {
#define RC_MODEL_TOTAL_BITS 11
/* Called twice: once at startup and once in rc_normalize() */
static void rc_read(rc_t *rc)
/* Called twice: once at startup (LZMA_FAST only) and once in rc_normalize() */
static size_inline void rc_read(rc_t *rc)
{
int buffer_size = safe_read(rc->fd, RC_BUFFER, RC_BUFFER_SIZE);
if (buffer_size <= 0)
@ -54,8 +55,17 @@ static void rc_read(rc_t *rc)
rc->buffer_end = RC_BUFFER + buffer_size;
}
/* Called twice, but one callsite is in speed_inline'd rc_is_bit_1() */
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++;
}
/* Called once */
static rc_t* rc_init(int fd) /*, int buffer_size) */
static ALWAYS_INLINE rc_t* rc_init(int fd) /*, int buffer_size) */
{
int i;
rc_t *rc;
@ -63,17 +73,18 @@ static rc_t* rc_init(int fd) /*, int buffer_size) */
rc = xmalloc(sizeof(*rc) + 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 ENABLE_FEATURE_LZMA_FAST
if (rc->ptr >= rc->buffer_end)
rc_read(rc);
rc->code = (rc->code << 8) | *rc->ptr++;
#else
rc_do_normalize(rc);
#endif
}
rc->range = 0xFFFFFFFF;
return rc;
}
@ -83,14 +94,6 @@ static ALWAYS_INLINE void rc_free(rc_t *rc)
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 ALWAYS_INLINE void rc_normalize(rc_t *rc)
{
if (rc->range < (1 << RC_TOP_BITS)) {
@ -98,49 +101,28 @@ static ALWAYS_INLINE void rc_normalize(rc_t *rc)
}
}
/* rc_is_bit_0 is called 9 times */
/* Why rc_is_bit_0_helper exists?
* Because we want to always expose (rc->code < rc->bound) to optimizer.
* Thus rc_is_bit_0 is always inlined, and rc_is_bit_0_helper is inlined
* only if we compile for speed.
*/
static speed_inline uint32_t rc_is_bit_0_helper(rc_t *rc, uint16_t *p)
/* rc_is_bit_1 is called 9 times */
static speed_inline int rc_is_bit_1(rc_t *rc, uint16_t *p)
{
rc_normalize(rc);
rc->bound = *p * (rc->range >> RC_MODEL_TOTAL_BITS);
return rc->bound;
}
static 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)
{
if (rc->code < rc->bound) {
rc->range = rc->bound;
*p += ((1 << RC_MODEL_TOTAL_BITS) - *p) >> RC_MOVE_BITS;
return 0;
}
rc->range -= rc->bound;
rc->code -= rc->bound;
*p -= *p >> RC_MOVE_BITS;
return 1;
}
/* Called 4 times in unlzma loop */
static int rc_get_bit(rc_t *rc, uint16_t *p, int *symbol)
static speed_inline 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;
}
int ret = rc_is_bit_1(rc, p);
*symbol = *symbol * 2 + ret;
return ret;
}
/* Called once */
@ -266,13 +248,13 @@ unpack_lzma_stream(int src_fd, int dst_fd)
header.dst_size = SWAP_LE64(header.dst_size);
if (header.dict_size == 0)
header.dict_size = 1;
header.dict_size++;
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));
num_probs += LZMA_LITERAL - LZMA_BASE_SIZE;
for (i = 0; i < num_probs; i++)
p[i] = (1 << RC_MODEL_TOTAL_BITS) >> 1;
@ -282,9 +264,8 @@ unpack_lzma_stream(int src_fd, int dst_fd)
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)) {
if (!rc_is_bit_1(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))
@ -340,27 +321,21 @@ unpack_lzma_stream(int src_fd, int dst_fd)
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);
if (!rc_is_bit_1(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 += LZMA_IS_REP_G0 - LZMA_IS_REP;
if (!rc_is_bit_1(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);
if (!rc_is_bit_1(rc, prob)) {
state = state < LZMA_NUM_LIT_STATES ? 9 : 11;
#if ENABLE_FEATURE_LZMA_FAST
pos = buffer_pos - rep0;
@ -372,25 +347,16 @@ unpack_lzma_stream(int src_fd, int dst_fd)
len = 1;
goto string;
#endif
} 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);
prob += LZMA_IS_REP_G1 - LZMA_IS_REP_G0;
distance = rep1;
if (rc_is_bit_1(rc, prob)) {
prob += LZMA_IS_REP_G2 - LZMA_IS_REP_G1;
distance = rep2;
if (rc_is_bit_1(rc, prob)) {
distance = rep3;
rep3 = rep2;
}
@ -404,24 +370,20 @@ unpack_lzma_stream(int src_fd, int dst_fd)
}
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));
if (!rc_is_bit_1(rc, prob_len)) {
prob_len += LZMA_LEN_LOW - LZMA_LEN_CHOICE
+ (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));
prob_len += LZMA_LEN_CHOICE_2 - LZMA_LEN_CHOICE;
if (!rc_is_bit_1(rc, prob_len)) {
prob_len += LZMA_LEN_MID - LZMA_LEN_CHOICE_2
+ (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;
prob_len += LZMA_LEN_HIGH - LZMA_LEN_CHOICE_2;
offset = ((1 << LZMA_LEN_NUM_LOW_BITS)
+ (1 << LZMA_LEN_NUM_MID_BITS));
num_bits = LZMA_LEN_NUM_HIGH_BITS;
@ -438,19 +400,20 @@ unpack_lzma_stream(int src_fd, int dst_fd)
((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);
rc_bit_tree_decode(rc, prob,
LZMA_NUM_POS_SLOT_BITS, &pos_slot);
rep0 = pos_slot;
if (pos_slot >= LZMA_START_POS_MODEL_INDEX) {
num_bits = (pos_slot >> 1) - 1;
rep0 = 2 | (pos_slot & 1);
prob = p + LZMA_ALIGN;
if (pos_slot < LZMA_END_POS_MODEL_INDEX) {
rep0 <<= num_bits;
prob = p + LZMA_SPEC_POS + rep0 - pos_slot - 1;
prob += LZMA_SPEC_POS - LZMA_ALIGN - 1 + rep0 - pos_slot;
} 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;
}
@ -461,8 +424,7 @@ unpack_lzma_stream(int src_fd, int dst_fd)
rep0 |= i;
i <<= 1;
}
} else
rep0 = pos_slot;
}
if (++rep0 == 0)
break;
}