6502/lzsa
1
0
Fork 0
mirror of https://github.com/emmanuel-marty/lzsa.git synced 2024-12-22 16:30:01 +00:00
Code Issues Projects Releases Wiki Activity

Added an option for unrolling long match copying

Browse Source 
Usually useless and costing +57 bytes, this option can bring dramatic performance improvements on very compressible data dominated by long matches
This commit is contained in:
introspec 2019-10-22 20:11:46 +01:00 committed by GitHub
parent e1e1276c96
commit d5d788946e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 49 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

59
asm/z80/unlzsa1_fast.asm
View File

@ -8,7 +8,8 @@
; ver.04 by spke (31/07/2019, small re-organization of macros);
; ver.05 by uniabis (22/08/2019, 107(-2) bytes, same speed);
; ver.06 by spke for LZSA 1.0.7 (27/08/2019, 111(+4) bytes, +2.1% speed);
; ver.07 by spke for LZSA 1.1.0 (25/09/2019, added full revision history)
; ver.07 by spke for LZSA 1.1.0 (25/09/2019, added full revision history);
; ver.08 by spke for LZSA 1.1.2 (22/10/2019, re-organized macros and added an option for unrolled copying of long matches)
;
; The data must be compressed using the command line compressor by Emmanuel Marty
; The compression is done as follows:
@ -56,6 +57,7 @@
; misrepresented as being the original software.
; 3. This notice may not be removed or altered from any source distribution.
; DEFINE UNROLL_LONG_MATCHES ; uncomment for faster decompression of very compressible data (+57 bytes)
; DEFINE BACKWARD_DECOMPRESS
IFNDEF BACKWARD_DECOMPRESS
@ -68,11 +70,11 @@
ex de,hl : add hl,de
ENDM
MACRO COPY_MATCH
ldi : ldi : ldir
MACRO COPY1
ldi
ENDM
MACRO BLOCKCOPY
MACRO COPYBC
ldir
ENDM
@ -87,11 +89,11 @@
ld a,d : sbc h : ld h,a ; 4*4+3*4 = 28t / 7 bytes
ENDM
MACRO COPY_MATCH
ldd : ldd : lddr
MACRO COPY1
ldd
ENDM
MACRO BLOCKCOPY
MACRO COPYBC
lddr
ENDM
@ -110,7 +112,8 @@ ShortOffset: ld d,#FF : add 3 : cp 15+3 : jr nc,LongerMatch
CopyMatch: ld c,a
.UseC NEXT_HL : ex (sp),hl ; BC = len, DE = offset, HL = dest, SP ->[dest,src]
ADD_OFFSET ; BC = len, DE = dest, HL = dest-offset, SP->[src]
COPY_MATCH : pop hl ; BC = 0, DE = dest, HL = src
COPY1 : COPY1 : COPYBC ; BC = 0, DE = dest
.popSrc pop hl ; HL = src
ReadToken: ; first a byte token "O|LLL|MMMM" is read from the stream,
; where LLL is the number of literals and MMMM is
@ -121,7 +124,7 @@ ReadToken: ; first a byte token "O|LLL|MMMM" is read from the stream,
rrca : rrca : rrca : rrca ; LLL<7 means 0..6 literals...
ld c,a : ld a,(hl)
NEXT_HL : BLOCKCOPY
NEXT_HL : COPYBC
; next we read the first byte of the offset
push de : ld e,(hl)
@ -132,6 +135,8 @@ LongOffset: ; read second byte of the offset
NEXT_HL : ld d,(hl)
add -128+3 : cp 15+3 : jp c,CopyMatch
IFNDEF UNROLL_LONG_MATCHES
; MMMM=15 indicates a multi-byte number of literals
LongerMatch: NEXT_HL : add (hl) : jr nc,CopyMatch
@ -146,12 +151,46 @@ LongerMatch: NEXT_HL : add (hl) : jr nc,CopyMatch
ld a,b : or c : jr nz,CopyMatch.UseC
pop de : ret
ELSE
; MMMM=15 indicates a multi-byte number of literals
LongerMatch: NEXT_HL : add (hl) : jr c,VeryLongMatch
ld c,a
.UseC NEXT_HL : ex (sp),hl
ADD_OFFSET
COPY1 : COPY1
; this is an unrolled equivalent of LDIR
xor a : sub c
and 16-1 : add a
ld (.jrOffset),a : jr nz,$+2
.jrOffset EQU $-1
.fastLDIR DUP 16
COPY1
EDUP
jp pe,.fastLDIR
jp CopyMatch.popSrc
VeryLongMatch: ; the codes are designed to overflow;
; the overflow value 1 means read 1 extra byte
; and overflow value 0 means read 2 extra bytes
.code1 ld b,a : NEXT_HL : ld c,(hl) : jr nz,LongerMatch.UseC
.code0 NEXT_HL : ld b,(hl)
; the two-byte match length equal to zero
; designates the end-of-data marker
ld a,b : or c : jr nz,LongerMatch.UseC
pop de : ret
ENDIF
MoreLiterals: ; there are three possible situations here
xor (hl) : exa
ld a,7 : NEXT_HL : add (hl) : jr c,ManyLiterals
CopyLiterals: ld c,a
.UseC NEXT_HL : BLOCKCOPY
.UseC NEXT_HL : COPYBC
push de : ld e,(hl)
exa : jp p,ShortOffset : jr LongOffset