SliTaz Repositories

     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/memtest64/stuff/unlzsa2.S	Thu Feb 16 14:18:05 2023 +0000
     1.3 @@ -0,0 +1,274 @@
     1.4 +// Lzsa2Decode:
     1.5 +#ifndef FLAT32
     1.6 +//   input   ds:si=inStream, es:di=outStream
     1.7 +//   output  outStream[], ds:si, es:di
     1.8 +	.code16
     1.9 +#define AX	%ax
    1.10 +#define BX	%bx
    1.11 +#define BP	%bp
    1.12 +#define SI	%si
    1.13 +#define DI	%di
    1.14 +#else
    1.15 +//   input   esi=inStream, edi=outStream
    1.16 +//   output  outStream[], ds:esi, es:edi
    1.17 +	.code32
    1.18 +#define AX	%eax
    1.19 +#define BX	%ebx
    1.20 +#define BP	%ebp
    1.21 +#define SI	%esi
    1.22 +#define DI	%edi
    1.23 +#endif
    1.24 +
    1.25 +MATCH_RUN_LEN		=	7
    1.26 +LITERALS_RUN_LEN	=	3
    1.27 +MIN_MATCH_SIZE		=	2
    1.28 +MIN_LITERALS_SIZE	=	0
    1.29 +
    1.30 +#define PACKED_ONLY			// assume no copy block, optional
    1.31 +//#define PARANOIA			// cover rare cases, optional
    1.32 + 
    1.33 +lzsa2main:
    1.34 +#ifdef PARANOIA
    1.35 +	cld
    1.36 +#endif
    1.37 +#ifdef FLAT16OUT
    1.38 +#define RAW_FORMAT
    1.39 +#endif
    1.40 +#ifndef RAW_FORMAT
    1.41 +# if defined(PARANOIA) && !defined(FLAT32) && !defined(FLAT16)
    1.42 +	xorw	%cx, %cx
    1.43 +	call	normalize
    1.44 +# endif
    1.45 +# ifndef NO_LZSA2_HEADER
    1.46 +	lodsw
    1.47 +	cmpw	$0x9E7B, %ax	// magic
    1.48 +	jne	lzsa2main
    1.49 +	lodsb
    1.50 +	testb	$0x20, %al	// lzsa2
    1.51 +	je	lzsa2main
    1.52 +# endif
    1.53 +lzsa2block:			// uncompress chunk
    1.54 +# if !defined(FLAT32) && !defined(FLAT16)
    1.55 +	xorw	%cx, %cx
    1.56 +	call	normalize
    1.57 +# endif
    1.58 +	lodsw			// block size
    1.59 +	xchgw	%ax, %cx
    1.60 +	lodsb
    1.61 +# ifndef PACKED_ONLY
    1.62 +	orb	%al, %al
    1.63 +	jns	lzsa2compressed
    1.64 +#  if !defined(FLAT32) && !defined(FLAT16OUT)
    1.65 +	movw	%cx, %dx
    1.66 +	movb	$0, %cl
    1.67 +	movb	$0, %dh
    1.68 +copytail:
    1.69 +	call	lzsa1movStr
    1.70 +	xchg	%dx, %cx
    1.71 +	incw	%cx
    1.72 +	loop	copytail
    1.73 +#  else
    1.74 +	movsb			// copy block
    1.75 +copylp:
    1.76 +	movsb			// copy block
    1.77 +	loop	copylp		// handle 64K case
    1.78 +#  endif
    1.79 +	jmp	lzsa2block	// end of block
    1.80 +lzsa2compressed:
    1.81 +	jne	lzsa2chunk	// 64Kb block
    1.82 +# endif
    1.83 +	jcxz	lzsa2quit	// bail if we hit EOD
    1.84 +# if !defined(FLAT16)
    1.85 +	pushw	%cx
    1.86 +	xorw	%cx, %cx
    1.87 +	call	normalize
    1.88 +# define NeedNormalize
    1.89 +	popw	%dx
    1.90 +	addw	%si, %dx
    1.91 +# else
    1.92 +	movw	%si, %dx
    1.93 +	addw	%cx, %dx
    1.94 +# endif
    1.95 +#else
    1.96 +# if !defined(FLAT16)
    1.97 +	xorw	%cx, %cx
    1.98 +	call	normalize
    1.99 +# define NeedNormalize
   1.100 +# endif
   1.101 +#endif
   1.102 +	movb	$0, %bh		// no nibble stored
   1.103 +lzsa2chunk:			// uncompress chunk
   1.104 +	lodsb			// get token XYZ|LL|MMM
   1.105 +	movb	%al, %bl	// keep token in bl
   1.106 +	movw	$LITERALS_RUN_LEN+256*MIN_LITERALS_SIZE, %cx
   1.107 +	shrb	%cl, %al	// shift literals length into place
   1.108 +	call	lzsa2len	// %cl = LITERALS_RUN_LEN
   1.109 +#if defined(NeedNormalize)
   1.110 +	pushw	%bp
   1.111 +	call	lzsa2movLit	// copy %cx literals from %ds:%si to %es:%di
   1.112 +	popw	%bp
   1.113 +#else
   1.114 +	rep	movsb		// copy %cx literals from %ds:%si to %es:%di
   1.115 +#endif
   1.116 +#ifndef RAW_FORMAT
   1.117 +maxsi:
   1.118 +	cmpw	%dx, %si
   1.119 +	jae	lzsa2block	// bail if we hit EOD
   1.120 +#endif
   1.121 +#ifdef FLAT32
   1.122 +	orl	$-1, %eax	// set offset bits 31-8 to 1
   1.123 +#else
   1.124 +	movb	$-1, %ah	// set offset bits 15-8 to 1
   1.125 +#endif
   1.126 +// XYZ
   1.127 +	testb	$0xC0, %bl	// check match offset mode in token (X bit)
   1.128 +	bt	$5, %bx		// move bit 5 to carry
   1.129 +	js	rep_match_or_large_offset
   1.130 +	jne	offset_9_bit
   1.131 +// 00Z 5-bit offset: read a nibble for offset bits 1-4 and use the inverted bit Z of the token as bit 0 of the offset.
   1.132 +//                   set bits 5-15 of the offset to 1.
   1.133 +	call	getByteFromNibbleAndC
   1.134 +	jmp	get_match_length
   1.135 +offset_9_bit:
   1.136 +// 01Z 9-bit offset: read a byte for offset bits 0-7 and use the inverted bit Z for bit 8 of the offset.
   1.137 +//                   set bits 9-15 of the offset to 1.
   1.138 +	sbbb	%cl, %ah	// clear bit 8 if Z bit is clear
   1.139 +	jmp	get_match_length_0
   1.140 +
   1.141 +getNibble:
   1.142 +	xorb	$0xF0, %bh	// toggle nibble stored flags
   1.143 +	movb	%bh, %al
   1.144 +	jns	gotnibble
   1.145 +	lodsb
   1.146 +	movb	$0xF0, %bh
   1.147 +	orb	%al, %bh
   1.148 +	shrb	$4, %al
   1.149 +gotnibble:
   1.150 +lzsa2quit:
   1.151 +	ret
   1.152 +
   1.153 +rep_match_or_large_offset:
   1.154 +	jpe	rep_match_or_16_bit
   1.155 +//10Z 13-bit offset: read a nibble for offset bits 9-12 and use the inverted bit Z for bit 8 of the offset,
   1.156 +//                   then read a byte for offset bits 0-7. set bits 13-15 of the offset to 1.
   1.157 +//                   substract 512 from the offset to get the final value.
   1.158 +	call	getByteFromNibbleAndC
   1.159 +	subb	$2, %al		// substract 512
   1.160 +	jmp	get_match_length_1
   1.161 +rep_match_or_16_bit:
   1.162 +	jc	repeat_match	// rep-match
   1.163 +//110 16-bit offset: read a byte for offset bits 8-15, then another byte for offset bits 0-7.
   1.164 +	lodsb			// Get 2-byte match offset
   1.165 +get_match_length_1:
   1.166 +	xchgb	%al, %ah
   1.167 +get_match_length_0:
   1.168 +	lodsb			// load match offset bits 0-7
   1.169 +get_match_length:
   1.170 +	xchgw	%ax, %bp	// bp: offset
   1.171 +repeat_match:
   1.172 +//111 repeat offset: reuse the offset value of the previous match command.
   1.173 +
   1.174 +	movb	%bl, %al	// %al: original token
   1.175 +	movw	$MATCH_RUN_LEN+256*MIN_MATCH_SIZE, %cx
   1.176 +	call	lzsa2len
   1.177 +#ifdef RAW_FORMAT
   1.178 +	jcxz	lzsa2quit	// bail if we hit EOD
   1.179 +#endif
   1.180 +#if !defined(FLAT32) && !defined(FLAT16OUT)
   1.181 +	pushw	%ds
   1.182 +	pushw	%si
   1.183 +	movw	%di, %si
   1.184 +	addw	%bp, %si
   1.185 +	movw	%es, %ax
   1.186 +	jc	axok
   1.187 +	subb	$0x10, %ah
   1.188 +axok:
   1.189 +.macro norm	reg
   1.190 +	movw	%si, \reg
   1.191 +	subw	%si, %dx
   1.192 +	andw	$0xF, %si
   1.193 +	addw	%si, %dx
   1.194 +	shrw	$4, \reg
   1.195 +	addw	\reg, %ax
   1.196 +	movw	%ax, %ds
   1.197 +	movw	%di, \reg
   1.198 +	andw	$0xF, %di
   1.199 +	shrw	$4, \reg
   1.200 +	movw	%es, %ax
   1.201 +	addw	\reg, %ax
   1.202 +	movw	%ax, %es
   1.203 +.endm 
   1.204 +	pushw	%bp
   1.205 +	pushw	%dx
   1.206 +# if defined(NeedNormalize) || defined(PARANOIA)
   1.207 +	call	lzsa2movStr	// copy string
   1.208 +# else
   1.209 +	norm	%bp
   1.210 +	rep movsb
   1.211 +# endif
   1.212 +	popw	%dx
   1.213 +	popw	%bp
   1.214 +	popw	%si
   1.215 +	popw	%ds
   1.216 +#else
   1.217 +	xchg	AX, SI		// save %si	
   1.218 +	lea	(BP,DI), SI
   1.219 +	rep movsb	%es:(SI), %es:(DI)
   1.220 +	xchg	AX, SI		// restore %si	
   1.221 +#endif
   1.222 +	jmp	lzsa2chunk
   1.223 +
   1.224 +getByteFromNibbleAndC:
   1.225 +	pushfw
   1.226 +	call	getNibble	// get nibble for offset bits 0-3
   1.227 +	popfw
   1.228 +	rclb	$1, %al
   1.229 +	xorb	$0xE1, %al	// set offset bits 7-5 to 1
   1.230 +	ret
   1.231 +
   1.232 +lzsa2len:			// get length in %cx
   1.233 +	andb	%cl, %al
   1.234 +	cbw			// clear %ah
   1.235 +	cmpb	%al, %cl
   1.236 +	jne	lzsa2minNumber	// S=0-2, L=0-6
   1.237 +	call	getNibble
   1.238 +	cmp	$0xF, %al
   1.239 +	jne	lzsa2noExtraByte
   1.240 +	addb	%al, %ch
   1.241 +	lodsb
   1.242 +lzsa2noExtraByte:
   1.243 +	addb	%cl, %ch
   1.244 +lzsa2minNumber:
   1.245 +	addb	%ch, %al
   1.246 +	jnc	lzsa2gotNumber  // 0-255
   1.247 +#if 0
   1.248 +	je	lzsa2BigNumber
   1.249 +	movb	%al, %ah	// S=256-767, L=256-1791
   1.250 +	lodsb	
   1.251 +	.byte	0xB1		// mask lodsw with movb $0xAD, %cl
   1.252 +lzsa2BigNumber:
   1.253 +#endif
   1.254 +	lodsw			// 0-65535
   1.255 +lzsa2gotNumber:
   1.256 +	xchgw	%ax, %cx
   1.257 +	ret
   1.258 +
   1.259 +#if defined(NeedNormalize) || defined(PARANOIA)
   1.260 +# if defined(PARANOIA)
   1.261 +lzsa2movlp:
   1.262 +	decw	%ch
   1.263 +	rep movsb
   1.264 +	incw	%ch
   1.265 +# endif
   1.266 +normalize:
   1.267 +lzsa2movLit:
   1.268 +	movw	%ds, %ax
   1.269 +lzsa2movStr:
   1.270 +	norm	%bp
   1.271 +# if defined(PARANOIA)
   1.272 +	cmpb	$0xFF, %ch	// catch FFFX case
   1.273 +	je	lzsa2movlp
   1.274 +# endif
   1.275 +	rep movsb
   1.276 +	ret
   1.277 +#endif