Apple-2-SW/dos33fsprogs
1
0
Fork 0
mirror of https://github.com/deater/dos33fsprogs.git synced 2024-11-01 01:06:33 +00:00
Code Issues Projects Releases Wiki Activity
master
dos33fsprogs/compression/lzsa_test/decompress_faster_v2.s

638 lines
20 KiB
ArmAsm
Raw Permalink Normal View History

lzsa: update benchmarks
2020-03-05 04:22:17 +00:00
; ***************************************************************************
; ***************************************************************************
;
; lzsa2_6502.s
;
; NMOS 6502 decompressor for data stored in Emmanuel Marty's LZSA2 format.
;
; This code is written for the ACME assembler.
;
; Optional code is presented for two minor 6502 optimizations that break
; compatibility with the current LZSA2 format standard.
;
; The code is 241 bytes for the small version, and 267 bytes for the normal.
;
; Copyright John Brandwood 2019.
;
; Distributed under the Boost Software License, Version 1.0.
; (See accompanying file LICENSE_1_0.txt or copy at
; http://www.boost.org/LICENSE_1_0.txt)
;
; ***************************************************************************
; ***************************************************************************
; ***************************************************************************
; ***************************************************************************
;
; Decompression Options & Macros
;
;
; Save 7 bytes of code, and 21 cycles every time that a
; 16-bit length is decoded?
;
; N.B. Setting this breaks compatibility with LZSA v1.2
;
LZSA_SWAP_LEN16 = 0
;
; Save 3 bytes of code, and 4 or 8 cycles when decoding
; an offset?
;
; N.B. Setting this breaks compatibility with LZSA v1.2
;
LZSA_SWAP_XZY = 0
;
; Choose size over space (within sane limits)?
;
LZSA_SMALL_SIZE = 0
;
; Remove code inlining to save space?
;
; This saves 15 bytes of code at the cost of 7% speed.
;
!if LZSA_SMALL_SIZE {
LZSA_NO_INLINE = 1
} else {
LZSA_NO_INLINE = 0
}
;
; Use smaller code for copying literals?
;
; This saves 11 bytes of code at the cost of 5% speed.
;
!if LZSA_SMALL_SIZE {
LZSA_SHORT_CP = 1
} else {
LZSA_SHORT_CP = 0
}
;
; Assume that we're decompressing from a large multi-bank
; compressed data file, and that the next bank may need to
; paged in when a page-boundary is crossed.
;
LZSA_FROM_BANK = 0
;
; We will read from or write to $FFFF. This prevents the
; use of the "INC ptrhi / BNE" trick and reduces speed.
;
LZSA_USE_FFFF = 0
;
; Macro to increment the source pointer to the next page.
;
!if LZSA_FROM_BANK {
!macro LZSA_INC_PAGE {
jsr lzsa2_next_page
}
} else {
!macro LZSA_INC_PAGE {
inc <lzsa_srcptr + 1
}
}
;
; Macro to read a byte from the compressed source data.
;
!if LZSA_NO_INLINE {
!macro LZSA_GET_SRC {
jsr lzsa2_get_byte
}
} else {
!macro LZSA_GET_SRC {
lda (lzsa_srcptr),y
inc <lzsa_srcptr + 0
bne .skip
+LZSA_INC_PAGE
.skip:
}
}
;
; Macro to speed up reading 50% of nibbles.
;
; This seems to save very few cycles compared to the
; increase in code size, and it isn't recommended.
;
LZSA_SLOW_NIBL = 1
!if (LZSA_SLOW_NIBL + LZSA_SMALL_SIZE) {
!macro LZSA_GET_NIBL {
jsr lzsa2_get_nibble ; Always call a function.
}
} else {
!macro LZSA_GET_NIBL {
lsr <lzsa_nibflg ; Is there a nibble waiting?
lda <lzsa_nibble ; Extract the lo-nibble.
bcs .skip
jsr lzsa2_new_nibble ; Extract the hi-nibble.
.skip: ora #$F0
}
}
; ***************************************************************************
; ***************************************************************************
;
; Data usage is last 11 bytes of zero-page.
;
lzsa_cmdbuf = $F5 ; 1 byte.
lzsa_nibflg = $F6 ; 1 byte.
lzsa_nibble = $F7 ; 1 byte.
lzsa_offset = $F8 ; 1 word.
lzsa_winptr = $FA ; 1 word.
lzsa_srcptr = $FC ; 1 word.
lzsa_dstptr = $FE ; 1 word.
lzsa_length = lzsa_winptr ; 1 word.
LZSA_SRC_LO = $FC
LZSA_SRC_HI = $FD
LZSA_DST_LO = $FE
LZSA_DST_HI = $FF
; ***************************************************************************
; ***************************************************************************
;
; lzsa2_unpack - Decompress data stored in Emmanuel Marty's LZSA2 format.
;
; Args: lzsa_srcptr = ptr to compessed data
; Args: lzsa_dstptr = ptr to output buffer
; Uses: lots!
;
; If compiled with LZSA_FROM_BANK, then lzsa_srcptr should be within the bank
; window range.
;
DECOMPRESS_LZSA2_FAST:
lzsa2_unpack: ldy #0 ; Initialize source index.
sty <lzsa_nibflg ; Initialize nibble buffer.
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) = 0 {
beq .cp_length ; always taken
.incsrc1:
inc <lzsa_srcptr + 1
bne .resume_src1 ; always taken
!if LZSA_SHORT_CP {
.incsrc2:
inc <lzsa_srcptr + 1
bne .resume_src2 ; always taken
.incdst:
inc <lzsa_dstptr + 1
bne .resume_dst ; always taken
}
}
;
; Copy bytes from compressed source data.
;
.cp_length: ldx #$00 ; Hi-byte of length or offset.
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) {
+LZSA_GET_SRC
} else {
lda (lzsa_srcptr),y
inc <lzsa_srcptr + 0
beq .incsrc1
}
.resume_src1:
sta <lzsa_cmdbuf ; Preserve this for later.
and #$18 ; Extract literal length.
beq .lz_offset ; Skip directly to match?
lsr ; Get 2-bit literal length.
lsr
lsr
cmp #$03 ; Extended length?
bne .got_cp_len
jsr .get_length ; X=0 table index for literals.
!if LZSA_SHORT_CP {
.got_cp_len: cmp #0 ; Check the lo-byte of length.
beq .put_cp_len
inx ; Increment # of pages to copy.
.put_cp_len: stx <lzsa_length
tax
.cp_page: lda (lzsa_srcptr),y
sta (lzsa_dstptr),y
inc <lzsa_srcptr + 0
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) {
bne .skip1
inc <lzsa_srcptr + 1
.skip1: inc <lzsa_dstptr + 0
bne .skip2
inc <lzsa_dstptr + 1
.skip2:
} else {
beq .incsrc2
.resume_src2:
inc <lzsa_dstptr + 0
beq .incdst
.resume_dst:
}
dex
bne .cp_page
dec <lzsa_length ; Any full pages left to copy?
bne .cp_page
} else {
.got_cp_len: tay ; Check the lo-byte of length.
beq .cp_page
inx ; Increment # of pages to copy.
.get_cp_src: clc ; Calc address of partial page.
adc <lzsa_srcptr + 0
sta <lzsa_srcptr + 0
bcs .get_cp_dst
dec <lzsa_srcptr + 1
.get_cp_dst: tya
clc ; Calc address of partial page.
adc <lzsa_dstptr + 0
sta <lzsa_dstptr + 0
bcs .get_cp_idx
dec <lzsa_dstptr + 1
.get_cp_idx: tya ; Negate the lo-byte of length.
eor #$FF
tay
iny
.cp_page: lda (lzsa_srcptr),y
sta (lzsa_dstptr),y
iny
bne .cp_page
inc <lzsa_srcptr + 1
inc <lzsa_dstptr + 1
dex ; Any full pages left to copy?
bne .cp_page
}
!if LZSA_SWAP_XZY {
;
; Shorter and faster path with NEW order of bits.
;
; STD NEW
; ================================
; xyz xzy
; 00z 0z0 5-bit offset
; 01z 0z1 9-bit offset
; 10z 1z0 13-bit offset
; 110 101 16-bit offset
; 111 111 repeat offset
; NVZ for a BIT instruction
;
; N.B. Saves 3 bytes in code length.
; get5 and get13 are 8 cycles faster.
; get9, get16, and rep are 4 cycles faster.
;
.lz_offset: lda #$20 ; Y bit in lzsa_cmdbuf.
bit <lzsa_cmdbuf
bmi .get_13_16_rep
bne .get_9_bits
.get_5_bits: dex ; X=$FF
.get_13_bits: +LZSA_GET_NIBL ; Always returns with CS.
bvc .get_5_skip
clc
.get_5_skip: rol ; Shift into position, set C.
cpx #$00 ; X=$FF for a 5-bit offset.
bne .set_offset
sbc #2 ; Subtract 512 because 13-bit
tax ; offset starts at $FE00.
bne .get_low8 ; Always NZ from previous TAX.
.get_9_bits: dex ; X=$FF if VC, X=$FE if VS.
bvc .get_low8
dex
bvs .get_low8 ; Always VS from previous BIT.
.get_13_16_rep: beq .get_13_bits ; Shares code with 5-bit path.
.get_16_rep: bvs .lz_length ; Repeat previous offset.
} else {
;
; Slower and longer path with STD order of bits.
;
; STD NEW
; ================================
; xyz xzy
; 00z 0z0 5-bit offset
; 01z 0z1 9-bit offset
; 10z 1z0 13-bit offset
; 110 101 16-bit offset
; 111 111 repeat offset
; NVZ for a BIT instruction
;
; N.B. Costs 3 bytes in code length.
; get5 and get13 are 8 cycles slower.
; get9, get16, and rep are 4 cycles slower.
;
.lz_offset: lda <lzsa_cmdbuf
asl
bcs .get_13_16_rep
asl
bcs .get_9_bits
.get_5_bits: dex ; X=$FF
.get_13_bits: asl
php
+LZSA_GET_NIBL ; Always returns with CS.
plp
rol ; Shift into position, set C.
eor #$01
cpx #$00 ; X=$FF for a 5-bit offset.
bne .set_offset
sbc #2 ; Subtract 512 because 13-bit
; offset starts at $FE00.
bne .get_low8x ; Always NZ from previous SBC.
.get_9_bits: dex ; X=$FF if CS, X=$FE if CC.
asl
bcc .get_low8
dex
bcs .get_low8 ; Always VS from previous BIT.
.get_13_16_rep: asl
bcc .get_13_bits ; Shares code with 5-bit path.
.get_16_rep: bmi .lz_length ; Repeat previous offset.
}
;
; Copy bytes from decompressed window.
;
; N.B. X=0 is expected and guaranteed when we get here.
;
.get_16_bits: jsr lzsa2_get_byte ; Get hi-byte of offset.
.get_low8x: tax
.get_low8:
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) {
+LZSA_GET_SRC ; Get lo-byte of offset.
} else {
lda (lzsa_srcptr),y
inc <lzsa_srcptr + 0
beq .incsrc3
.resume_src3:
}
.set_offset: stx <lzsa_offset + 1 ; Save new offset.
sta <lzsa_offset + 0
.lz_length: ldx #$00 ; Hi-byte of length.
lda <lzsa_cmdbuf
and #$07
clc
adc #$02
cmp #$09 ; Extended length?
bne .got_lz_len
inx
jsr .get_length ; X=1 table index for match.
.got_lz_len: eor #$FF ; Negate the lo-byte of length
tay ; and check for zero.
iny
beq .calc_lz_addr
eor #$FF
inx ; Increment # of pages to copy.
clc ; Calc destination for partial
adc <lzsa_dstptr + 0 ; page.
sta <lzsa_dstptr + 0
bcs .calc_lz_addr
dec <lzsa_dstptr + 1
.calc_lz_addr: clc ; Calc address of match.
lda <lzsa_dstptr + 0 ; N.B. Offset is negative!
adc <lzsa_offset + 0
sta <lzsa_winptr + 0
lda <lzsa_dstptr + 1
adc <lzsa_offset + 1
sta <lzsa_winptr + 1
.lz_page: lda (lzsa_winptr),y
sta (lzsa_dstptr),y
iny
bne .lz_page
inc <lzsa_winptr + 1
inc <lzsa_dstptr + 1
dex ; Any full pages left to copy?
bne .lz_page
jmp .cp_length ; Loop around to the beginning.
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) = 0 {
.incsrc3:
inc <lzsa_srcptr + 1
bne .resume_src3 ; always taken
}
;
; Lookup tables to differentiate literal and match lengths.
;
.nibl_len_tbl: !byte 3 + $10 ; 0+3 (for literal).
!byte 9 + $10 ; 2+7 (for match).
.byte_len_tbl: !byte 18 - 1 ; 0+3+15 - CS (for literal).
!byte 24 - 1 ; 2+7+15 - CS (for match).
;
; Get 16-bit length in X:A register pair.
;
.get_length: +LZSA_GET_NIBL
cmp #$FF ; Extended length?
bcs .byte_length
adc .nibl_len_tbl,x ; Always CC from previous CMP.
.got_length: ldx #$00 ; Set hi-byte of 4 & 8 bit
rts ; lengths.
.byte_length: jsr lzsa2_get_byte ; So rare, this can be slow!
adc .byte_len_tbl,x ; Always CS from previous CMP.
bcc .got_length
beq .finished
!if LZSA_SWAP_LEN16 {
.word_length: jsr lzsa2_get_byte ; So rare, this can be slow!
tax
} else {
.word_length: jsr lzsa2_get_byte ; So rare, this can be slow!
pha
jsr lzsa2_get_byte ; So rare, this can be slow!
tax
pla
rts
}
lzsa2_get_byte:
lda (lzsa_srcptr),y ; Subroutine version for when
inc <lzsa_srcptr + 0 ; inlining isn't advantageous.
beq lzsa2_next_page
rts
lzsa2_next_page:
inc <lzsa_srcptr + 1 ; Inc & test for bank overflow.
!if LZSA_FROM_BANK {
bmi lzsa2_next_bank ; Change for target hardware!
}
rts
.finished: pla ; Decompression completed, pop
pla ; return address.
rts
;
; Get a nibble value from compressed data in A.
;
!if (LZSA_SLOW_NIBL | LZSA_SMALL_SIZE) {
lzsa2_get_nibble:
lsr <lzsa_nibflg ; Is there a nibble waiting?
lda <lzsa_nibble ; Extract the lo-nibble.
bcs .got_nibble
inc <lzsa_nibflg ; Reset the flag.
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) {
+LZSA_GET_SRC
} else {
lda (lzsa_srcptr),y
inc <lzsa_srcptr + 0
beq .incsrc4
.resume_src4:
}
sta <lzsa_nibble ; Preserve for next time.
lsr ; Extract the hi-nibble.
lsr
lsr
lsr
!if LZSA_SWAP_XZY {
sec ; Offset code relies on CS.
}
.got_nibble: ora #$F0
rts
} else {
lzsa2_new_nibble:
inc <lzsa_nibflg ; Reset the flag.
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) {
+LZSA_GET_SRC
} else {
lda (lzsa_srcptr),y
inc <lzsa_srcptr + 0
beq .incsrc4
.resume_src4:
}
sta <lzsa_nibble ; Preserve for next time.
lsr ; Extract the hi-nibble.
lsr
lsr
lsr
!if LZSA_SWAP_XZY {
sec ; Offset code relies on CS.
}
rts
}
!if (LZSA_FROM_BANK | LZSA_NO_INLINE | LZSA_USE_FFFF) = 0 {
.incsrc4:
inc <lzsa_srcptr + 1
bne .resume_src4 ; always taken
}
Reference in New Issue Copy Permalink