| 1 |
; inffasx64.asm is a hand tuned assembler version of inffast.c - fast decoding |
| 2 |
; version for AMD64 on Windows using Microsoft C compiler |
| 3 |
; |
| 4 |
; inffasx64.asm is automatically convert from AMD64 portion of inffas86.c |
| 5 |
; inffasx64.asm is called by inffas8664.c, which contain more info. |
| 6 |
|
| 7 |
|
| 8 |
; to compile this file, I use option |
| 9 |
; ml64.exe /Flinffasx64 /c /Zi inffasx64.asm |
| 10 |
; with Microsoft Macro Assembler (x64) for AMD64 |
| 11 |
; |
| 12 |
|
| 13 |
; This file compile with Microsoft Macro Assembler (x64) for AMD64 |
| 14 |
; |
| 15 |
; ml64.exe is given with Visual Studio 2005/2008/2010 and Windows WDK |
| 16 |
; |
| 17 |
; (you can get Windows WDK with ml64 for AMD64 from |
| 18 |
; http://www.microsoft.com/whdc/Devtools/wdk/default.mspx for low price) |
| 19 |
; |
| 20 |
|
| 21 |
|
| 22 |
.code |
| 23 |
inffas8664fnc PROC |
| 24 |
|
| 25 |
; see http://weblogs.asp.net/oldnewthing/archive/2004/01/14/58579.aspx and |
| 26 |
; http://msdn.microsoft.com/library/en-us/kmarch/hh/kmarch/64bitAMD_8e951dd2-ee77-4728-8702-55ce4b5dd24a.xml.asp |
| 27 |
; |
| 28 |
; All registers must be preserved across the call, except for |
| 29 |
; rax, rcx, rdx, r8, r-9, r10, and r11, which are scratch. |
| 30 |
|
| 31 |
|
| 32 |
mov [rsp-8],rsi |
| 33 |
mov [rsp-16],rdi |
| 34 |
mov [rsp-24],r12 |
| 35 |
mov [rsp-32],r13 |
| 36 |
mov [rsp-40],r14 |
| 37 |
mov [rsp-48],r15 |
| 38 |
mov [rsp-56],rbx |
| 39 |
|
| 40 |
mov rax,rcx |
| 41 |
|
| 42 |
mov [rax+8], rbp ; /* save regs rbp and rsp */ |
| 43 |
mov [rax], rsp |
| 44 |
|
| 45 |
mov rsp, rax ; /* make rsp point to &ar */ |
| 46 |
|
| 47 |
mov rsi, [rsp+16] ; /* rsi = in */ |
| 48 |
mov rdi, [rsp+32] ; /* rdi = out */ |
| 49 |
mov r9, [rsp+24] ; /* r9 = last */ |
| 50 |
mov r10, [rsp+48] ; /* r10 = end */ |
| 51 |
mov rbp, [rsp+64] ; /* rbp = lcode */ |
| 52 |
mov r11, [rsp+72] ; /* r11 = dcode */ |
| 53 |
mov rdx, [rsp+80] ; /* rdx = hold */ |
| 54 |
mov ebx, [rsp+88] ; /* ebx = bits */ |
| 55 |
mov r12d, [rsp+100] ; /* r12d = lmask */ |
| 56 |
mov r13d, [rsp+104] ; /* r13d = dmask */ |
| 57 |
; /* r14d = len */ |
| 58 |
; /* r15d = dist */ |
| 59 |
|
| 60 |
|
| 61 |
cld |
| 62 |
cmp r10, rdi |
| 63 |
je L_one_time ; /* if only one decode left */ |
| 64 |
cmp r9, rsi |
| 65 |
|
| 66 |
jne L_do_loop |
| 67 |
|
| 68 |
|
| 69 |
L_one_time: |
| 70 |
mov r8, r12 ; /* r8 = lmask */ |
| 71 |
cmp bl, 32 |
| 72 |
ja L_get_length_code_one_time |
| 73 |
|
| 74 |
lodsd ; /* eax = *(uint *)in++ */ |
| 75 |
mov cl, bl ; /* cl = bits, needs it for shifting */ |
| 76 |
add bl, 32 ; /* bits += 32 */ |
| 77 |
shl rax, cl |
| 78 |
or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ |
| 79 |
jmp L_get_length_code_one_time |
| 80 |
|
| 81 |
ALIGN 4 |
| 82 |
L_while_test: |
| 83 |
cmp r10, rdi |
| 84 |
jbe L_break_loop |
| 85 |
cmp r9, rsi |
| 86 |
jbe L_break_loop |
| 87 |
|
| 88 |
L_do_loop: |
| 89 |
mov r8, r12 ; /* r8 = lmask */ |
| 90 |
cmp bl, 32 |
| 91 |
ja L_get_length_code ; /* if (32 < bits) */ |
| 92 |
|
| 93 |
lodsd ; /* eax = *(uint *)in++ */ |
| 94 |
mov cl, bl ; /* cl = bits, needs it for shifting */ |
| 95 |
add bl, 32 ; /* bits += 32 */ |
| 96 |
shl rax, cl |
| 97 |
or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ |
| 98 |
|
| 99 |
L_get_length_code: |
| 100 |
and r8, rdx ; /* r8 &= hold */ |
| 101 |
mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */ |
| 102 |
|
| 103 |
mov cl, ah ; /* cl = this.bits */ |
| 104 |
sub bl, ah ; /* bits -= this.bits */ |
| 105 |
shr rdx, cl ; /* hold >>= this.bits */ |
| 106 |
|
| 107 |
test al, al |
| 108 |
jnz L_test_for_length_base ; /* if (op != 0) 45.7% */ |
| 109 |
|
| 110 |
mov r8, r12 ; /* r8 = lmask */ |
| 111 |
shr eax, 16 ; /* output this.val char */ |
| 112 |
stosb |
| 113 |
|
| 114 |
L_get_length_code_one_time: |
| 115 |
and r8, rdx ; /* r8 &= hold */ |
| 116 |
mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */ |
| 117 |
|
| 118 |
L_dolen: |
| 119 |
mov cl, ah ; /* cl = this.bits */ |
| 120 |
sub bl, ah ; /* bits -= this.bits */ |
| 121 |
shr rdx, cl ; /* hold >>= this.bits */ |
| 122 |
|
| 123 |
test al, al |
| 124 |
jnz L_test_for_length_base ; /* if (op != 0) 45.7% */ |
| 125 |
|
| 126 |
shr eax, 16 ; /* output this.val char */ |
| 127 |
stosb |
| 128 |
jmp L_while_test |
| 129 |
|
| 130 |
ALIGN 4 |
| 131 |
L_test_for_length_base: |
| 132 |
mov r14d, eax ; /* len = this */ |
| 133 |
shr r14d, 16 ; /* len = this.val */ |
| 134 |
mov cl, al |
| 135 |
|
| 136 |
test al, 16 |
| 137 |
jz L_test_for_second_level_length ; /* if ((op & 16) == 0) 8% */ |
| 138 |
and cl, 15 ; /* op &= 15 */ |
| 139 |
jz L_decode_distance ; /* if (!op) */ |
| 140 |
|
| 141 |
L_add_bits_to_len: |
| 142 |
sub bl, cl |
| 143 |
xor eax, eax |
| 144 |
inc eax |
| 145 |
shl eax, cl |
| 146 |
dec eax |
| 147 |
and eax, edx ; /* eax &= hold */ |
| 148 |
shr rdx, cl |
| 149 |
add r14d, eax ; /* len += hold & mask[op] */ |
| 150 |
|
| 151 |
L_decode_distance: |
| 152 |
mov r8, r13 ; /* r8 = dmask */ |
| 153 |
cmp bl, 32 |
| 154 |
ja L_get_distance_code ; /* if (32 < bits) */ |
| 155 |
|
| 156 |
lodsd ; /* eax = *(uint *)in++ */ |
| 157 |
mov cl, bl ; /* cl = bits, needs it for shifting */ |
| 158 |
add bl, 32 ; /* bits += 32 */ |
| 159 |
shl rax, cl |
| 160 |
or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ |
| 161 |
|
| 162 |
L_get_distance_code: |
| 163 |
and r8, rdx ; /* r8 &= hold */ |
| 164 |
mov eax, [r11+r8*4] ; /* eax = dcode[hold & dmask] */ |
| 165 |
|
| 166 |
L_dodist: |
| 167 |
mov r15d, eax ; /* dist = this */ |
| 168 |
shr r15d, 16 ; /* dist = this.val */ |
| 169 |
mov cl, ah |
| 170 |
sub bl, ah ; /* bits -= this.bits */ |
| 171 |
shr rdx, cl ; /* hold >>= this.bits */ |
| 172 |
mov cl, al ; /* cl = this.op */ |
| 173 |
|
| 174 |
test al, 16 ; /* if ((op & 16) == 0) */ |
| 175 |
jz L_test_for_second_level_dist |
| 176 |
and cl, 15 ; /* op &= 15 */ |
| 177 |
jz L_check_dist_one |
| 178 |
|
| 179 |
L_add_bits_to_dist: |
| 180 |
sub bl, cl |
| 181 |
xor eax, eax |
| 182 |
inc eax |
| 183 |
shl eax, cl |
| 184 |
dec eax ; /* (1 << op) - 1 */ |
| 185 |
and eax, edx ; /* eax &= hold */ |
| 186 |
shr rdx, cl |
| 187 |
add r15d, eax ; /* dist += hold & ((1 << op) - 1) */ |
| 188 |
|
| 189 |
L_check_window: |
| 190 |
mov r8, rsi ; /* save in so from can use it's reg */ |
| 191 |
mov rax, rdi |
| 192 |
sub rax, [rsp+40] ; /* nbytes = out - beg */ |
| 193 |
|
| 194 |
cmp eax, r15d |
| 195 |
jb L_clip_window ; /* if (dist > nbytes) 4.2% */ |
| 196 |
|
| 197 |
mov ecx, r14d ; /* ecx = len */ |
| 198 |
mov rsi, rdi |
| 199 |
sub rsi, r15 ; /* from = out - dist */ |
| 200 |
|
| 201 |
sar ecx, 1 |
| 202 |
jnc L_copy_two ; /* if len % 2 == 0 */ |
| 203 |
|
| 204 |
rep movsw |
| 205 |
mov al, [rsi] |
| 206 |
mov [rdi], al |
| 207 |
inc rdi |
| 208 |
|
| 209 |
mov rsi, r8 ; /* move in back to %rsi, toss from */ |
| 210 |
jmp L_while_test |
| 211 |
|
| 212 |
L_copy_two: |
| 213 |
rep movsw |
| 214 |
mov rsi, r8 ; /* move in back to %rsi, toss from */ |
| 215 |
jmp L_while_test |
| 216 |
|
| 217 |
ALIGN 4 |
| 218 |
L_check_dist_one: |
| 219 |
cmp r15d, 1 ; /* if dist 1, is a memset */ |
| 220 |
jne L_check_window |
| 221 |
cmp [rsp+40], rdi ; /* if out == beg, outside window */ |
| 222 |
je L_check_window |
| 223 |
|
| 224 |
mov ecx, r14d ; /* ecx = len */ |
| 225 |
mov al, [rdi-1] |
| 226 |
mov ah, al |
| 227 |
|
| 228 |
sar ecx, 1 |
| 229 |
jnc L_set_two |
| 230 |
mov [rdi], al |
| 231 |
inc rdi |
| 232 |
|
| 233 |
L_set_two: |
| 234 |
rep stosw |
| 235 |
jmp L_while_test |
| 236 |
|
| 237 |
ALIGN 4 |
| 238 |
L_test_for_second_level_length: |
| 239 |
test al, 64 |
| 240 |
jnz L_test_for_end_of_block ; /* if ((op & 64) != 0) */ |
| 241 |
|
| 242 |
xor eax, eax |
| 243 |
inc eax |
| 244 |
shl eax, cl |
| 245 |
dec eax |
| 246 |
and eax, edx ; /* eax &= hold */ |
| 247 |
add eax, r14d ; /* eax += len */ |
| 248 |
mov eax, [rbp+rax*4] ; /* eax = lcode[val+(hold&mask[op])]*/ |
| 249 |
jmp L_dolen |
| 250 |
|
| 251 |
ALIGN 4 |
| 252 |
L_test_for_second_level_dist: |
| 253 |
test al, 64 |
| 254 |
jnz L_invalid_distance_code ; /* if ((op & 64) != 0) */ |
| 255 |
|
| 256 |
xor eax, eax |
| 257 |
inc eax |
| 258 |
shl eax, cl |
| 259 |
dec eax |
| 260 |
and eax, edx ; /* eax &= hold */ |
| 261 |
add eax, r15d ; /* eax += dist */ |
| 262 |
mov eax, [r11+rax*4] ; /* eax = dcode[val+(hold&mask[op])]*/ |
| 263 |
jmp L_dodist |
| 264 |
|
| 265 |
ALIGN 4 |
| 266 |
L_clip_window: |
| 267 |
mov ecx, eax ; /* ecx = nbytes */ |
| 268 |
mov eax, [rsp+92] ; /* eax = wsize, prepare for dist cmp */ |
| 269 |
neg ecx ; /* nbytes = -nbytes */ |
| 270 |
|
| 271 |
cmp eax, r15d |
| 272 |
jb L_invalid_distance_too_far ; /* if (dist > wsize) */ |
| 273 |
|
| 274 |
add ecx, r15d ; /* nbytes = dist - nbytes */ |
| 275 |
cmp dword ptr [rsp+96], 0 |
| 276 |
jne L_wrap_around_window ; /* if (write != 0) */ |
| 277 |
|
| 278 |
mov rsi, [rsp+56] ; /* from = window */ |
| 279 |
sub eax, ecx ; /* eax -= nbytes */ |
| 280 |
add rsi, rax ; /* from += wsize - nbytes */ |
| 281 |
|
| 282 |
mov eax, r14d ; /* eax = len */ |
| 283 |
cmp r14d, ecx |
| 284 |
jbe L_do_copy ; /* if (nbytes >= len) */ |
| 285 |
|
| 286 |
sub eax, ecx ; /* eax -= nbytes */ |
| 287 |
rep movsb |
| 288 |
mov rsi, rdi |
| 289 |
sub rsi, r15 ; /* from = &out[ -dist ] */ |
| 290 |
jmp L_do_copy |
| 291 |
|
| 292 |
ALIGN 4 |
| 293 |
L_wrap_around_window: |
| 294 |
mov eax, [rsp+96] ; /* eax = write */ |
| 295 |
cmp ecx, eax |
| 296 |
jbe L_contiguous_in_window ; /* if (write >= nbytes) */ |
| 297 |
|
| 298 |
mov esi, [rsp+92] ; /* from = wsize */ |
| 299 |
add rsi, [rsp+56] ; /* from += window */ |
| 300 |
add rsi, rax ; /* from += write */ |
| 301 |
sub rsi, rcx ; /* from -= nbytes */ |
| 302 |
sub ecx, eax ; /* nbytes -= write */ |
| 303 |
|
| 304 |
mov eax, r14d ; /* eax = len */ |
| 305 |
cmp eax, ecx |
| 306 |
jbe L_do_copy ; /* if (nbytes >= len) */ |
| 307 |
|
| 308 |
sub eax, ecx ; /* len -= nbytes */ |
| 309 |
rep movsb |
| 310 |
mov rsi, [rsp+56] ; /* from = window */ |
| 311 |
mov ecx, [rsp+96] ; /* nbytes = write */ |
| 312 |
cmp eax, ecx |
| 313 |
jbe L_do_copy ; /* if (nbytes >= len) */ |
| 314 |
|
| 315 |
sub eax, ecx ; /* len -= nbytes */ |
| 316 |
rep movsb |
| 317 |
mov rsi, rdi |
| 318 |
sub rsi, r15 ; /* from = out - dist */ |
| 319 |
jmp L_do_copy |
| 320 |
|
| 321 |
ALIGN 4 |
| 322 |
L_contiguous_in_window: |
| 323 |
mov rsi, [rsp+56] ; /* rsi = window */ |
| 324 |
add rsi, rax |
| 325 |
sub rsi, rcx ; /* from += write - nbytes */ |
| 326 |
|
| 327 |
mov eax, r14d ; /* eax = len */ |
| 328 |
cmp eax, ecx |
| 329 |
jbe L_do_copy ; /* if (nbytes >= len) */ |
| 330 |
|
| 331 |
sub eax, ecx ; /* len -= nbytes */ |
| 332 |
rep movsb |
| 333 |
mov rsi, rdi |
| 334 |
sub rsi, r15 ; /* from = out - dist */ |
| 335 |
jmp L_do_copy ; /* if (nbytes >= len) */ |
| 336 |
|
| 337 |
ALIGN 4 |
| 338 |
L_do_copy: |
| 339 |
mov ecx, eax ; /* ecx = len */ |
| 340 |
rep movsb |
| 341 |
|
| 342 |
mov rsi, r8 ; /* move in back to %esi, toss from */ |
| 343 |
jmp L_while_test |
| 344 |
|
| 345 |
L_test_for_end_of_block: |
| 346 |
test al, 32 |
| 347 |
jz L_invalid_literal_length_code |
| 348 |
mov dword ptr [rsp+116], 1 |
| 349 |
jmp L_break_loop_with_status |
| 350 |
|
| 351 |
L_invalid_literal_length_code: |
| 352 |
mov dword ptr [rsp+116], 2 |
| 353 |
jmp L_break_loop_with_status |
| 354 |
|
| 355 |
L_invalid_distance_code: |
| 356 |
mov dword ptr [rsp+116], 3 |
| 357 |
jmp L_break_loop_with_status |
| 358 |
|
| 359 |
L_invalid_distance_too_far: |
| 360 |
mov dword ptr [rsp+116], 4 |
| 361 |
jmp L_break_loop_with_status |
| 362 |
|
| 363 |
L_break_loop: |
| 364 |
mov dword ptr [rsp+116], 0 |
| 365 |
|
| 366 |
L_break_loop_with_status: |
| 367 |
; /* put in, out, bits, and hold back into ar and pop esp */ |
| 368 |
mov [rsp+16], rsi ; /* in */ |
| 369 |
mov [rsp+32], rdi ; /* out */ |
| 370 |
mov [rsp+88], ebx ; /* bits */ |
| 371 |
mov [rsp+80], rdx ; /* hold */ |
| 372 |
|
| 373 |
mov rax, [rsp] ; /* restore rbp and rsp */ |
| 374 |
mov rbp, [rsp+8] |
| 375 |
mov rsp, rax |
| 376 |
|
| 377 |
|
| 378 |
|
| 379 |
mov rsi,[rsp-8] |
| 380 |
mov rdi,[rsp-16] |
| 381 |
mov r12,[rsp-24] |
| 382 |
mov r13,[rsp-32] |
| 383 |
mov r14,[rsp-40] |
| 384 |
mov r15,[rsp-48] |
| 385 |
mov rbx,[rsp-56] |
| 386 |
|
| 387 |
ret 0 |
| 388 |
; : |
| 389 |
; : "m" (ar) |
| 390 |
; : "memory", "%rax", "%rbx", "%rcx", "%rdx", "%rsi", "%rdi", |
| 391 |
; "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15" |
| 392 |
; ); |
| 393 |
|
| 394 |
inffas8664fnc ENDP |
| 395 |
;_TEXT ENDS |
| 396 |
END |