generic-library/isa-l
1
0
Fork 0
mirror of https://github.com/intel/isa-l.git synced 2025-03-04 19:13:28 +01:00
Code Issues Projects Releases Wiki Activity

igzip: Setup for variable lookback distance

Browse Source 
Change-Id: Idd52c9392113dfc54feea3c66916a7f5aa128bef
Signed-off-by: Roy Oursler <roy.j.oursler@intel.com>
This commit is contained in:
Roy Oursler 2018-06-06 18:48:40 -04:00
parent f421ea8d7a
commit 6317ce2b78
12 changed files with 77 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
igzip/data_struct2.asm
View File

@ -159,6 +159,7 @@ START_FIELDS ;; isal_zstate
FIELD _total_in_start,4, 4 FIELD _total_in_start,4, 4
FIELD _block_next, 4, 4 FIELD _block_next, 4, 4
FIELD _block_end, 4, 4 FIELD _block_end, 4, 4
FIELD _dist_mask, 4, 4
FIELD _bitbuf, _BitBuf2_size, _BitBuf2_align FIELD _bitbuf, _BitBuf2_size, _BitBuf2_align
FIELD _crc, 4, 4 FIELD _crc, 4, 4
FIELD _state, 4, 4 FIELD _state, 4, 4
@ -203,7 +204,8 @@ FIELD _level_buf_size, 4, 4
FIELD _level_buf, 8, 8 FIELD _level_buf, 8, 8
FIELD _end_of_stream, 2, 2 FIELD _end_of_stream, 2, 2
FIELD _flush, 2, 2 FIELD _flush, 2, 2
FIELD _gzip_flag, 4, 4 FIELD _gzip_flag, 2, 2
FIELD _hist_bits, 2, 2
FIELD _internal_state, _isal_zstate_size, _isal_zstate_align FIELD _internal_state, _isal_zstate_size, _isal_zstate_align
%assign _isal_zstream_size _FIELD_OFFSET %assign _isal_zstream_size _FIELD_OFFSET
@ -215,6 +217,7 @@ _internal_state_block_end equ _internal_state+_block_end
_internal_state_b_bytes_valid equ _internal_state+_b_bytes_valid _internal_state_b_bytes_valid equ _internal_state+_b_bytes_valid
_internal_state_b_bytes_processed equ _internal_state+_b_bytes_processed _internal_state_b_bytes_processed equ _internal_state+_b_bytes_processed
_internal_state_crc equ _internal_state+_crc _internal_state_crc equ _internal_state+_crc
_internal_state_dist_mask equ _internal_state+_dist_mask
_internal_state_bitbuf equ _internal_state+_bitbuf _internal_state_bitbuf equ _internal_state+_bitbuf
_internal_state_state equ _internal_state+_state _internal_state_state equ _internal_state+_state
_internal_state_count equ _internal_state+_count _internal_state_count equ _internal_state+_count

19
igzip/igzip.c
View File

@ -911,6 +911,18 @@ static inline void reset_match_history(struct isal_zstream *stream)
} }
} }
static void inline set_dist_mask(struct isal_zstream *stream)
{
struct isal_zstate *state = &stream->internal_state;
uint32_t hist_size = (1 << (stream->hist_bits));
if (stream->hist_bits != 0 && hist_size < IGZIP_HIST_SIZE)
state->dist_mask = hist_size - 1;
else
state->dist_mask = IGZIP_HIST_SIZE - 1;
}
void isal_deflate_init(struct isal_zstream *stream) void isal_deflate_init(struct isal_zstream *stream)
{ {
struct isal_zstate *state = &stream->internal_state; struct isal_zstate *state = &stream->internal_state;
@ -924,6 +936,7 @@ void isal_deflate_init(struct isal_zstream *stream)
stream->end_of_stream = 0; stream->end_of_stream = 0;
stream->flush = NO_FLUSH; stream->flush = NO_FLUSH;
stream->gzip_flag = 0; stream->gzip_flag = 0;
stream->hist_bits = 0;
state->block_next = 0; state->block_next = 0;
state->block_end = 0; state->block_end = 0;
@ -1097,6 +1110,7 @@ int isal_deflate_stateless(struct isal_zstream *stream)
state->state = ZSTATE_NEW_HDR; state->state = ZSTATE_NEW_HDR;
state->crc = 0; state->crc = 0;
state->has_level_buf_init = 0; state->has_level_buf_init = 0;
set_dist_mask(stream);
if (stream->flush == NO_FLUSH) if (stream->flush == NO_FLUSH)
stream->end_of_stream = 1; stream->end_of_stream = 1;
@ -1248,13 +1262,16 @@ int isal_deflate(struct isal_zstream *stream)
hist_size = get_hist_size(stream, start_in, buf_hist_start); hist_size = get_hist_size(stream, start_in, buf_hist_start);
if (state->has_hist == IGZIP_NO_HIST) { if (state->has_hist == IGZIP_NO_HIST) {
set_dist_mask(stream);
stream->total_in -= buffered_size; stream->total_in -= buffered_size;
reset_match_history(stream); reset_match_history(stream);
stream->total_in += buffered_size; stream->total_in += buffered_size;
buf_hist_start = state->b_bytes_processed; buf_hist_start = state->b_bytes_processed;
} else if (state->has_hist == IGZIP_DICT_HIST) } else if (state->has_hist == IGZIP_DICT_HIST) {
set_dist_mask(stream);
isal_deflate_hash(stream, state->buffer, state->b_bytes_processed); isal_deflate_hash(stream, state->buffer, state->b_bytes_processed);
}
in_size = stream->avail_in + buffered_size; in_size = stream->avail_in + buffered_size;
out_size = stream->total_out; out_size = stream->total_out;

6
igzip/igzip_base.c
View File

@ -36,6 +36,7 @@ void isal_deflate_body_base(struct isal_zstream *stream)
struct isal_zstate *state = &stream->internal_state; struct isal_zstate *state = &stream->internal_state;
uint16_t *last_seen = state->head; uint16_t *last_seen = state->head;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
if (stream->avail_in == 0) { if (stream->avail_in == 0) {
if (stream->end_of_stream || stream->flush != NO_FLUSH) if (stream->end_of_stream || stream->flush != NO_FLUSH)
@ -62,7 +63,7 @@ void isal_deflate_body_base(struct isal_zstream *stream)
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
/* The -1 are to handle the case when dist = 0 */ /* The -1 are to handle the case when dist = 0 */
if (dist - 1 < IGZIP_HIST_SIZE - 1) { if (dist - 1 < hist_size) {
assert(dist != 0); assert(dist != 0);
match_length = compare258(next_in - dist, next_in, 258); match_length = compare258(next_in - dist, next_in, 258);
@ -122,6 +123,7 @@ void isal_deflate_finish_base(struct isal_zstream *stream)
struct isal_zstate *state = &stream->internal_state; struct isal_zstate *state = &stream->internal_state;
uint16_t *last_seen = state->head; uint16_t *last_seen = state->head;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
set_buf(&state->bitbuf, stream->next_out, stream->avail_out); set_buf(&state->bitbuf, stream->next_out, stream->avail_out);
@ -141,7 +143,7 @@ void isal_deflate_finish_base(struct isal_zstream *stream)
dist = (next_in - file_start - last_seen[hash]) & 0xFFFF; dist = (next_in - file_start - last_seen[hash]) & 0xFFFF;
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
if (dist - 1 < IGZIP_HIST_SIZE - 1) { /* The -1 are to handle the case when dist = 0 */ if (dist - 1 < hist_size) { /* The -1 are to handle the case when dist = 0 */
match_length = match_length =
compare258(next_in - dist, next_in, end_in - next_in); compare258(next_in - dist, next_in, end_in - next_in);

7
igzip/igzip_body.asm
View File

@ -224,13 +224,14 @@ isal_deflate_body_ %+ ARCH %+ :
align 16 align 16
.loop2: .loop2:
mov tmp3 %+ d, dword [stream + _internal_state_dist_mask]
; if (state->bitbuf.is_full()) { ; if (state->bitbuf.is_full()) {
cmp m_out_buf, [stream + _internal_state_bitbuf_m_out_end] cmp m_out_buf, [stream + _internal_state_bitbuf_m_out_end]
ja .output_end ja .output_end
xor dist, dist xor dist, dist
xor dist2, dist2 xor dist2, dist2
xor tmp3, tmp3
lea tmp1, [file_start + f_i] lea tmp1, [file_start + f_i]
@ -252,13 +253,13 @@ isal_deflate_body_ %+ ARCH %+ :
mov [stream + _internal_state_head + 2 * hash2], f_i %+ w mov [stream + _internal_state_head + 2 * hash2], f_i %+ w
; if ((dist-1) < (D-1)) { ; if ((dist-1) < (D-1)) {
and dist %+ d, (D-1) and dist, tmp3
neg dist neg dist
shr tmp8, 8 shr tmp8, 8
compute_hash tmp2, tmp8 compute_hash tmp2, tmp8
and dist2 %+ d, (D-1) and dist2, tmp3
neg dist2 neg dist2
;; Check for long len/dist match (>7) with first literal ;; Check for long len/dist match (>7) with first literal

4
igzip/igzip_finish.asm
View File

@ -141,6 +141,8 @@ skip_SLOP:
skip_write_first_byte: skip_write_first_byte:
loop2: loop2:
mov tmp3 %+ d, dword [stream + _internal_state_dist_mask]
; if (state->bitbuf.is_full()) { ; if (state->bitbuf.is_full()) {
cmp m_out_buf, [stream + _internal_state_bitbuf_m_out_end] cmp m_out_buf, [stream + _internal_state_bitbuf_m_out_end]
ja end_loop_2 ja end_loop_2
@ -164,7 +166,7 @@ loop2:
; if ((dist-1) <= (D-1)) { ; if ((dist-1) <= (D-1)) {
mov tmp1 %+ d, dist %+ d mov tmp1 %+ d, dist %+ d
sub tmp1 %+ d, 1 sub tmp1 %+ d, 1
cmp tmp1 %+ d, (D-1) cmp tmp1 %+ d, tmp3 %+ d
jae encode_literal jae encode_literal
; len = f_end_i - f_i; ; len = f_end_i - f_i;

18
igzip/igzip_gen_icf_map_lh1_04.asm
View File

@ -101,6 +101,7 @@
%ifidn __OUTPUT_FORMAT__, win64 %ifidn __OUTPUT_FORMAT__, win64
%define stack_size 10*16 + 6 * 8 + 8 %define stack_size 10*16 + 6 * 8 + 8
%define local_storage_offset (stack_size - 8)
%define func(x) proc_frame x %define func(x) proc_frame x
%macro FUNC_SAVE 0 %macro FUNC_SAVE 0
@ -143,20 +144,27 @@
add rsp, stack_size add rsp, stack_size
%endm %endm
%else %else
%define stack_size 8
%define local_storage_offset 0
%define func(x) x: %define func(x) x:
%macro FUNC_SAVE 0 %macro FUNC_SAVE 0
push rbp push rbp
push r12 push r12
push r13 push r13
sub rsp, stack_size
%endm %endm
%macro FUNC_RESTORE 0 %macro FUNC_RESTORE 0
add rsp, stack_size
pop r13 pop r13
pop r12 pop r12
pop rbp pop rbp
%endm %endm
%endif %endif
%define dist_mask_offset local_storage_offset
%define VECT_SIZE 8 %define VECT_SIZE 8
%define HASH_BYTES 2 %define HASH_BYTES 2
@ -174,6 +182,8 @@ func(gen_icf_map_lh1_04)
jge end_main jge end_main
;; Prep for main loop ;; Prep for main loop
mov tmp %+ d, dword [stream + _internal_state_dist_mask]
mov [rsp + dist_mask_offset], tmp
mov tmp, stream mov tmp, stream
mov level_buf, [stream + _level_buf] mov level_buf, [stream + _level_buf]
sub f_i_end, LA sub f_i_end, LA
@ -207,7 +217,7 @@ func(gen_icf_map_lh1_04)
dec tmp dec tmp
sub tmp %+ w, word [hash_table + HASH_BYTES * hash] sub tmp %+ w, word [hash_table + HASH_BYTES * hash]
and tmp %+ d, [dist_mask] and tmp %+ d, [rsp + dist_mask_offset]
neg tmp neg tmp
;; Check first 8 bytes of match ;; Check first 8 bytes of match
@ -307,7 +317,7 @@ loop1:
lea next_in, [f_i + file_start] lea next_in, [f_i + file_start]
;; Calculate look back dists ;; Calculate look back dists
vpbroadcastd ydist_mask, [dist_mask] vpbroadcastd ydist_mask, [rsp + dist_mask_offset]
vpaddd ydists, ydists_lookup, yones vpaddd ydists, ydists_lookup, yones
vpsubd ydists, yindex, ydists vpsubd ydists, yindex, ydists
vpand ydists, ydists, ydist_mask vpand ydists, ydists, ydist_mask
@ -503,7 +513,7 @@ loop1_end:
lea next_in, [f_i + file_start] lea next_in, [f_i + file_start]
;; Calculate look back dists ;; Calculate look back dists
vpbroadcastd ydist_mask, [dist_mask] vpbroadcastd ydist_mask, [rsp + dist_mask_offset]
vpaddd ydists, ydists_lookup, yones vpaddd ydists, ydists_lookup, yones
vpsubd ydists, yindex, ydists vpsubd ydists, yindex, ydists
vpand ydists, ydists, ydist_mask vpand ydists, ydists, ydist_mask
@ -710,8 +720,6 @@ ones:
%define PROD2 0x97B1 %define PROD2 0x97B1
hash_prod: hash_prod:
dw PROD1, PROD2 dw PROD1, PROD2
dist_mask:
dd D-1
null_dist_syms: null_dist_syms:
dd LIT dd LIT
hash_mask: hash_mask:

4
igzip/igzip_gen_icf_map_lh1_06.asm
View File

@ -174,6 +174,7 @@ func(gen_icf_map_lh1_06)
jge end_main jge end_main
;; Prep for main loop ;; Prep for main loop
vpbroadcastd zdist_mask, dword [stream + _internal_state_dist_mask]
mov tmp, stream mov tmp, stream
mov level_buf, [stream + _level_buf] mov level_buf, [stream + _level_buf]
sub f_i_end, LA sub f_i_end, LA
@ -189,7 +190,6 @@ func(gen_icf_map_lh1_06)
vbroadcasti32x4 zbswap, [bswap_shuf] vbroadcasti32x4 zbswap, [bswap_shuf]
vpbroadcastd zthirty, [thirty] vpbroadcastd zthirty, [thirty]
vmovdqu64 zrot_left, [drot_left] vmovdqu64 zrot_left, [drot_left]
vpbroadcastd zdist_mask, dword [dist_mask]
vpbroadcastd zshortest_matches, [shortest_matches] vpbroadcastd zshortest_matches, [shortest_matches]
vpbroadcastd ztwofiftyfour, [twofiftyfour] vpbroadcastd ztwofiftyfour, [twofiftyfour]
vpbroadcastd znull_dist_syms, [null_dist_syms] vpbroadcastd znull_dist_syms, [null_dist_syms]
@ -569,8 +569,6 @@ thirty:
dd 0x1e dd 0x1e
twofiftyfour: twofiftyfour:
dd 0xfe dd 0xfe
dist_mask:
dd D-1
hash_mask: hash_mask:
dd HASH_MAP_HASH_MASK dd HASH_MAP_HASH_MASK
lit_len_mask: lit_len_mask:

15
igzip/igzip_icf_base.c
View File

@ -44,6 +44,7 @@ void isal_deflate_icf_body_hash8k_base(struct isal_zstream *stream)
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *last_seen = level_buf->hash8k.hash_table; uint16_t *last_seen = level_buf->hash8k.hash_table;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
if (stream->avail_in == 0) { if (stream->avail_in == 0) {
if (stream->end_of_stream || stream->flush != NO_FLUSH) if (stream->end_of_stream || stream->flush != NO_FLUSH)
@ -76,7 +77,7 @@ void isal_deflate_icf_body_hash8k_base(struct isal_zstream *stream)
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
/* The -1 are to handle the case when dist = 0 */ /* The -1 are to handle the case when dist = 0 */
if (dist - 1 < IGZIP_HIST_SIZE - 1) { if (dist - 1 < hist_size) {
assert(dist != 0); assert(dist != 0);
match_length = compare258(next_in - dist, next_in, 258); match_length = compare258(next_in - dist, next_in, 258);
@ -139,6 +140,7 @@ void isal_deflate_icf_body_hash_hist_base(struct isal_zstream *stream)
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *last_seen = level_buf->hash_hist.hash_table; uint16_t *last_seen = level_buf->hash_hist.hash_table;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
if (stream->avail_in == 0) { if (stream->avail_in == 0) {
if (stream->end_of_stream || stream->flush != NO_FLUSH) if (stream->end_of_stream || stream->flush != NO_FLUSH)
@ -171,7 +173,7 @@ void isal_deflate_icf_body_hash_hist_base(struct isal_zstream *stream)
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
/* The -1 are to handle the case when dist = 0 */ /* The -1 are to handle the case when dist = 0 */
if (dist - 1 < IGZIP_HIST_SIZE - 1) { if (dist - 1 < hist_size) {
assert(dist != 0); assert(dist != 0);
match_length = compare258(next_in - dist, next_in, 258); match_length = compare258(next_in - dist, next_in, 258);
@ -234,6 +236,7 @@ void isal_deflate_icf_finish_hash8k_base(struct isal_zstream *stream)
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *last_seen = level_buf->hash8k.hash_table; uint16_t *last_seen = level_buf->hash8k.hash_table;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
start_in = stream->next_in; start_in = stream->next_in;
end_in = start_in + stream->avail_in; end_in = start_in + stream->avail_in;
@ -263,7 +266,7 @@ void isal_deflate_icf_finish_hash8k_base(struct isal_zstream *stream)
dist = (next_in - file_start - last_seen[hash]) & 0xFFFF; dist = (next_in - file_start - last_seen[hash]) & 0xFFFF;
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
if (dist - 1 < IGZIP_HIST_SIZE - 1) { /* The -1 are to handle the case when dist = 0 */ if (dist - 1 < hist_size) { /* The -1 are to handle the case when dist = 0 */
match_length = compare258(next_in - dist, next_in, end_in - next_in); match_length = compare258(next_in - dist, next_in, end_in - next_in);
if (match_length >= SHORTEST_MATCH) { if (match_length >= SHORTEST_MATCH) {
@ -343,6 +346,7 @@ void isal_deflate_icf_finish_hash_hist_base(struct isal_zstream *stream)
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *last_seen = level_buf->hash_hist.hash_table; uint16_t *last_seen = level_buf->hash_hist.hash_table;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
start_in = stream->next_in; start_in = stream->next_in;
end_in = start_in + stream->avail_in; end_in = start_in + stream->avail_in;
@ -372,7 +376,7 @@ void isal_deflate_icf_finish_hash_hist_base(struct isal_zstream *stream)
dist = (next_in - file_start - last_seen[hash]) & 0xFFFF; dist = (next_in - file_start - last_seen[hash]) & 0xFFFF;
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
if (dist - 1 < IGZIP_HIST_SIZE - 1) { /* The -1 are to handle the case when dist = 0 */ if (dist - 1 < hist_size) { /* The -1 are to handle the case when dist = 0 */
match_length = compare258(next_in - dist, next_in, end_in - next_in); match_length = compare258(next_in - dist, next_in, end_in - next_in);
if (match_length >= SHORTEST_MATCH) { if (match_length >= SHORTEST_MATCH) {
@ -452,6 +456,7 @@ void isal_deflate_icf_finish_hash_map_base(struct isal_zstream *stream)
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *last_seen = level_buf->hash_map.hash_table; uint16_t *last_seen = level_buf->hash_map.hash_table;
uint8_t *file_start = stream->next_in - stream->total_in; uint8_t *file_start = stream->next_in - stream->total_in;
uint32_t hist_size = state->dist_mask;
start_in = stream->next_in; start_in = stream->next_in;
end_in = start_in + stream->avail_in; end_in = start_in + stream->avail_in;
@ -480,7 +485,7 @@ void isal_deflate_icf_finish_hash_map_base(struct isal_zstream *stream)
dist = (next_in - file_start - last_seen[hash]) & 0xFFFF; dist = (next_in - file_start - last_seen[hash]) & 0xFFFF;
last_seen[hash] = (uint64_t) (next_in - file_start); last_seen[hash] = (uint64_t) (next_in - file_start);
if (dist - 1 < IGZIP_HIST_SIZE - 1) { /* The -1 are to handle the case when dist = 0 */ if (dist - 1 < hist_size) { /* The -1 are to handle the case when dist = 0 */
match_length = compare258(next_in - dist, next_in, end_in - next_in); match_length = compare258(next_in - dist, next_in, end_in - next_in);
if (match_length >= SHORTEST_MATCH) { if (match_length >= SHORTEST_MATCH) {

3
igzip/igzip_icf_body.c
View File

@ -77,6 +77,7 @@ uint64_t gen_icf_map_h1_base(struct isal_zstream *stream,
uint64_t match; uint64_t match;
struct level_buf *level_buf = (struct level_buf *)stream->level_buf; struct level_buf *level_buf = (struct level_buf *)stream->level_buf;
uint16_t *hash_table = level_buf->hash_map.hash_table; uint16_t *hash_table = level_buf->hash_map.hash_table;
uint32_t hist_size = stream->internal_state.dist_mask;
if (input_size < ISAL_LOOK_AHEAD) if (input_size < ISAL_LOOK_AHEAD)
return 0; return 0;
@ -97,7 +98,7 @@ uint64_t gen_icf_map_h1_base(struct isal_zstream *stream,
while (next_in < end_in - ISAL_LOOK_AHEAD) { while (next_in < end_in - ISAL_LOOK_AHEAD) {
hash = compute_hash(*(uint32_t *) next_in) & HASH_MAP_HASH_MASK; hash = compute_hash(*(uint32_t *) next_in) & HASH_MAP_HASH_MASK;
dist = (next_in - file_start - hash_table[hash]); dist = (next_in - file_start - hash_table[hash]);
dist = ((dist - 1) & (IGZIP_HIST_SIZE - 1)) + 1; dist = ((dist - 1) & hist_size) + 1;
hash_table[hash] = (uint64_t) (next_in - file_start); hash_table[hash] = (uint64_t) (next_in - file_start);
match_bytes = *(uint64_t *) (next_in - dist); match_bytes = *(uint64_t *) (next_in - dist);

17
igzip/igzip_icf_body_h1_gr_bt.asm
View File

@ -103,11 +103,12 @@ global %1
m_out_end equ 0 ; local variable (8 bytes) m_out_end equ 0 ; local variable (8 bytes)
m_out_start equ 8 m_out_start equ 8
f_end_i_mem_offset equ 16 dist_mask_offset equ 16
stream_offset equ 24 f_end_i_mem_offset equ 24
gpr_save_mem_offset equ 32 ; gpr save area (8*8 bytes) stream_offset equ 32
gpr_save_mem_offset equ 40 ; gpr save area (8*8 bytes)
xmm_save_mem_offset equ gpr_save_mem_offset + 8*8 ; xmm save area (4*16 bytes) (16 byte aligned) xmm_save_mem_offset equ gpr_save_mem_offset + 8*8 ; xmm save area (4*16 bytes) (16 byte aligned)
stack_size equ 5*8 + 8*8 + 4*16 stack_size equ 7*8 + 8*8 + 4*16
;;; 8 because stack address is odd multiple of 8 after a function call and ;;; 8 because stack address is odd multiple of 8 after a function call and
;;; we want it aligned to 16 bytes ;;; we want it aligned to 16 bytes
@ -182,6 +183,8 @@ isal_deflate_icf_body_ %+ METHOD %+ _ %+ ARCH %+ :
mov [rsp + stream_offset], stream mov [rsp + stream_offset], stream
mov byte [stream + _internal_state_has_eob], 0 mov byte [stream + _internal_state_has_eob], 0
mov tmp1 %+ d, dword[stream + _internal_state_dist_mask]
mov [rsp + dist_mask_offset], tmp1
; state->bitbuf.set_buf(stream->next_out, stream->avail_out); ; state->bitbuf.set_buf(stream->next_out, stream->avail_out);
mov level_buf, [stream + _level_buf] mov level_buf, [stream + _level_buf]
@ -230,13 +233,13 @@ isal_deflate_icf_body_ %+ METHOD %+ _ %+ ARCH %+ :
align 16 align 16
.loop2: .loop2:
mov tmp3 %+ d, [rsp + dist_mask_offset]
; if (state->bitbuf.is_full()) { ; if (state->bitbuf.is_full()) {
cmp m_out_buf, [rsp + m_out_end] cmp m_out_buf, [rsp + m_out_end]
ja .output_end ja .output_end
xor dist, dist xor dist, dist
xor dist2, dist2 xor dist2, dist2
xor tmp3, tmp3
lea tmp1, [file_start + f_i] lea tmp1, [file_start + f_i]
@ -258,14 +261,14 @@ isal_deflate_icf_body_ %+ METHOD %+ _ %+ ARCH %+ :
mov [hash_table + 2 * hash2], f_i %+ w mov [hash_table + 2 * hash2], f_i %+ w
; if ((dist-1) < (D-1)) { ; if ((dist-1) < (D-1)) {
and dist %+ d, (D-1) and dist %+ d, tmp3 %+ d
neg dist neg dist
shr tmp2, 24 shr tmp2, 24
compute_hash hash2, tmp2 compute_hash hash2, tmp2
and hash2 %+ d, HASH_MASK and hash2 %+ d, HASH_MASK
and dist2 %+ d, (D-1) and dist2 %+ d, tmp3 %+ d
neg dist2 neg dist2
;; Check for long len/dist match (>7) with first literal ;; Check for long len/dist match (>7) with first literal

6
igzip/igzip_icf_finish.asm
View File

@ -86,6 +86,7 @@
f_end_i_mem_offset equ 0 ; local variable (8 bytes) f_end_i_mem_offset equ 0 ; local variable (8 bytes)
m_out_end equ 8 m_out_end equ 8
m_out_start equ 16 m_out_start equ 16
dist_mask_offset equ 24
stack_size equ 32 stack_size equ 32
%xdefine HASH_MASK HASH8K_HASH_MASK %xdefine HASH_MASK HASH8K_HASH_MASK
@ -108,6 +109,7 @@ isal_deflate_icf_finish_ %+ METHOD %+ _01:
%endif %endif
; state->bitbuf.set_buf(stream->next_out, stream->avail_out); ; state->bitbuf.set_buf(stream->next_out, stream->avail_out);
mov tmp2, [stream + _internal_state_dist_mask]
mov level_buf, [stream + _level_buf] mov level_buf, [stream + _level_buf]
mov m_out_buf, [level_buf + _icf_buf_next] mov m_out_buf, [level_buf + _icf_buf_next]
mov [rsp + m_out_start], m_out_buf mov [rsp + m_out_start], m_out_buf
@ -115,6 +117,7 @@ isal_deflate_icf_finish_ %+ METHOD %+ _01:
add tmp1, m_out_buf add tmp1, m_out_buf
sub tmp1, 4 sub tmp1, 4
mov [rsp + dist_mask_offset], tmp2
mov [rsp + m_out_end], tmp1 mov [rsp + m_out_end], tmp1
mov hufftables, [stream + _hufftables] mov hufftables, [stream + _hufftables]
@ -150,6 +153,7 @@ isal_deflate_icf_finish_ %+ METHOD %+ _01:
.skip_write_first_byte: .skip_write_first_byte:
.loop2: .loop2:
mov tmp3 %+ d, [rsp + dist_mask_offset]
; if (state->bitbuf.is_full()) { ; if (state->bitbuf.is_full()) {
cmp m_out_buf, [rsp + m_out_end] cmp m_out_buf, [rsp + m_out_end]
ja .end_loop_2 ja .end_loop_2
@ -173,7 +177,7 @@ isal_deflate_icf_finish_ %+ METHOD %+ _01:
; if ((dist-1) <= (D-1)) { ; if ((dist-1) <= (D-1)) {
mov tmp1 %+ d, dist %+ d mov tmp1 %+ d, dist %+ d
sub tmp1 %+ d, 1 sub tmp1 %+ d, 1
cmp tmp1 %+ d, (D-1) cmp tmp1 %+ d, tmp3 %+ d
jae .encode_literal jae .encode_literal
; len = f_end_i - f_i; ; len = f_end_i - f_i;

5
include/igzip_lib.h
View File

@ -314,6 +314,7 @@ struct isal_zstate {
uint32_t total_in_start; //!< Not used, may be replaced with something else uint32_t total_in_start; //!< Not used, may be replaced with something else
uint32_t block_next; //!< Start of current deflate block in the input uint32_t block_next; //!< Start of current deflate block in the input
uint32_t block_end; //!< End of current deflate block in the input uint32_t block_end; //!< End of current deflate block in the input
uint32_t dist_mask; //!< Distance mask used.
struct BitBuf2 bitbuf; //!< Bit Buffer struct BitBuf2 bitbuf; //!< Bit Buffer
uint32_t crc; //!< Current crc uint32_t crc; //!< Current crc
enum isal_zstate_state state; //!< Current state in processing the data stream enum isal_zstate_state state; //!< Current state in processing the data stream
@ -365,8 +366,8 @@ struct isal_zstream {
uint8_t * level_buf; //!< User allocated buffer required for different compression levels uint8_t * level_buf; //!< User allocated buffer required for different compression levels
uint16_t end_of_stream; //!< non-zero if this is the last input buffer uint16_t end_of_stream; //!< non-zero if this is the last input buffer
uint16_t flush; //!< Flush type can be NO_FLUSH, SYNC_FLUSH or FULL_FLUSH uint16_t flush; //!< Flush type can be NO_FLUSH, SYNC_FLUSH or FULL_FLUSH
uint32_t gzip_flag; //!< Indicate if gzip compression is to be performed uint16_t gzip_flag; //!< Indicate if gzip compression is to be performed
uint16_t hist_bits; //!< Log base 2 of maximum lookback distance, 0 is use default
struct isal_zstate internal_state; //!< Internal state for this stream struct isal_zstate internal_state; //!< Internal state for this stream
}; };