gvmat64.asm 14.7 KB
Newer Older
M
Mark Adler 已提交
1 2 3 4 5 6
;uInt longest_match_x64(
;    deflate_state *s,
;    IPos cur_match);                             /* current match */

; gvmat64.asm -- Asm portion of the optimized longest_match for 32 bits x86
; Copyright (C) 1995-2005 Jean-loup Gailly, Brian Raiter and Gilles Vollant.
M
Mark Adler 已提交
7 8 9 10 11 12 13
;
; File written by Gilles Vollant, by converting to assembly the longest_match
;  from Jean-loup Gailly in deflate.c of zLib and infoZip zip.
;
;  and by taking inspiration on asm686 with masm, optimised assembly code
;        from Brian Raiter, written 1998
;
M
Mark Adler 已提交
14 15 16 17
;         http://www.zlib.net
;         http://www.winimage.com/zLibDll
;         http://www.muppetlabs.com/~breadbox/software/assembly.html
;
M
Mark Adler 已提交
18 19 20 21
; to compile this file for infozip Zip, I use option:
;   ml64.exe /Flgvmat64 /c /Zi /DINFOZIP gvmat64.asm
;
; to compile this file for zLib, I use option:
M
Mark Adler 已提交
22
;   ml64.exe /Flgvmat64 /c /Zi gvmat64.asm
M
Mark Adler 已提交
23 24 25
; Be carrefull to adapt zlib1222add below to your version of zLib
;   (if you use a version of zLib before 1.0.4 or after 1.2.2.2, change
;    value of zlib1222add later)
M
Mark Adler 已提交
26
;
M
Mark Adler 已提交
27 28 29
; This file compile with Microsoft Macro Assembler (x64) for AMD64
;
;   ml64.exe is given with Visual Studio 2005 and Windows 2003 server DDK
M
Mark Adler 已提交
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
;
;   (you can get Windows 2003 server DDK with ml64 and cl for AMD64 from
;      http://www.microsoft.com/whdc/devtools/ddk/default.mspx for low price)
;


;uInt longest_match(s, cur_match)
;    deflate_state *s;
;    IPos cur_match;                             /* current match */
.code
longest_match PROC


;LocalVarsSize   equ 88
 LocalVarsSize   equ 72

; register used : rax,rbx,rcx,rdx,rsi,rdi,r8,r9,r10,r11,r12
; free register :  r14,r15
; register can be saved : rsp

 chainlenwmask   equ  rsp + 8 - LocalVarsSize    ; high word: current chain len
                                                 ; low word: s->wmask
;window          equ  rsp + xx - LocalVarsSize   ; local copy of s->window ; stored in r10
;windowbestlen   equ  rsp + xx - LocalVarsSize   ; s->window + bestlen , use r10+r11
;scanstart       equ  rsp + xx - LocalVarsSize   ; first two bytes of string ; stored in r12w
;scanend         equ  rsp + xx - LocalVarsSize   ; last two bytes of string use ebx
;scanalign       equ  rsp + xx - LocalVarsSize   ; dword-misalignment of string r13
;bestlen         equ  rsp + xx - LocalVarsSize   ; size of best match so far -> r11d
;scan            equ  rsp + xx - LocalVarsSize   ; ptr to string wanting match -> r9
M
Mark Adler 已提交
59 60 61 62
IFDEF INFOZIP
ELSE
 nicematch       equ  (rsp + 16 - LocalVarsSize) ; a good enough match size
ENDIF
M
Mark Adler 已提交
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99

save_rdi        equ  rsp + 24 - LocalVarsSize
save_rsi        equ  rsp + 32 - LocalVarsSize
save_rbx        equ  rsp + 40 - LocalVarsSize
save_rbp        equ  rsp + 48 - LocalVarsSize
save_r12        equ  rsp + 56 - LocalVarsSize
save_r13        equ  rsp + 64 - LocalVarsSize
;save_r14        equ  rsp + 72 - LocalVarsSize
;save_r15        equ  rsp + 80 - LocalVarsSize



;  all the +4 offsets are due to the addition of pending_buf_size (in zlib
;  in the deflate_state structure since the asm code was first written
;  (if you compile with zlib 1.0.4 or older, remove the +4).
;  Note : these value are good with a 8 bytes boundary pack structure


    MAX_MATCH           equ     258
    MIN_MATCH           equ     3
    MIN_LOOKAHEAD       equ     (MAX_MATCH+MIN_MATCH+1)


;;; Offsets for fields in the deflate_state structure. These numbers
;;; are calculated from the definition of deflate_state, with the
;;; assumption that the compiler will dword-align the fields. (Thus,
;;; changing the definition of deflate_state could easily cause this
;;; program to crash horribly, without so much as a warning at
;;; compile time. Sigh.)

;  all the +zlib1222add offsets are due to the addition of fields
;  in zlib in the deflate_state structure since the asm code was first written
;  (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)").
;  (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0").
;  if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8").


M
Mark Adler 已提交
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
IFDEF INFOZIP

_DATA   SEGMENT
COMM    window_size:DWORD
; WMask ; 7fff
COMM    window:BYTE:010040H
COMM    prev:WORD:08000H
; MatchLen : unused
; PrevMatch : unused
COMM    strstart:DWORD
COMM    match_start:DWORD
; Lookahead : ignore
COMM    prev_length:DWORD ; PrevLen
COMM    max_chain_length:DWORD
COMM    good_match:DWORD
COMM    nice_match:DWORD
prev_ad equ OFFSET prev
window_ad equ OFFSET window
nicematch equ nice_match
_DATA ENDS
WMask equ 07fffh

ELSE

  IFNDEF zlib1222add
    zlib1222add equ 8
  ENDIF
M
Mark Adler 已提交
127 128 129 130 131 132 133 134 135 136 137 138 139 140
dsWSize         equ 56+zlib1222add+(zlib1222add/2)
dsWMask         equ 64+zlib1222add+(zlib1222add/2)
dsWindow        equ 72+zlib1222add
dsPrev          equ 88+zlib1222add
dsMatchLen      equ 128+zlib1222add
dsPrevMatch     equ 132+zlib1222add
dsStrStart      equ 140+zlib1222add
dsMatchStart    equ 144+zlib1222add
dsLookahead     equ 148+zlib1222add
dsPrevLen       equ 152+zlib1222add
dsMaxChainLen   equ 156+zlib1222add
dsGoodMatch     equ 172+zlib1222add
dsNiceMatch     equ 176+zlib1222add

M
Mark Adler 已提交
141 142 143 144 145 146 147 148 149 150 151 152
window_size     equ [ rcx + dsWSize]
WMask           equ [ rcx + dsWMask]
window_ad       equ [ rcx + dsWindow]
prev_ad         equ [ rcx + dsPrev]
strstart        equ [ rcx + dsStrStart]
match_start     equ [ rcx + dsMatchStart]
Lookahead       equ [ rcx + dsLookahead] ; 0ffffffffh on infozip
prev_length     equ [ rcx + dsPrevLen]
max_chain_length equ [ rcx + dsMaxChainLen]
good_match      equ [ rcx + dsGoodMatch]
nice_match      equ [ rcx + dsNiceMatch]
ENDIF
M
Mark Adler 已提交
153 154 155 156 157 158 159

; parameter 1 in r8(deflate state s), param 2 in rdx (cur match)

; see http://weblogs.asp.net/oldnewthing/archive/2004/01/14/58579.aspx and
; http://msdn.microsoft.com/library/en-us/kmarch/hh/kmarch/64bitAMD_8e951dd2-ee77-4728-8702-55ce4b5dd24a.xml.asp
;
; All registers must be preserved across the call, except for
M
Mark Adler 已提交
160
;   rax, rcx, rdx, r8, r9, r10, and r11, which are scratch.
M
Mark Adler 已提交
161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180



;;; Save registers that the compiler may be using, and adjust esp to
;;; make room for our stack frame.


;;; Retrieve the function arguments. r8d will hold cur_match
;;; throughout the entire function. edx will hold the pointer to the
;;; deflate_state structure during the function's setup (before
;;; entering the main loop.

; parameter 1 in rcx (deflate_state* s), param 2 in edx -> r8 (cur match)

; this clear high 32 bits of r8, which can be garbage in both r8 and rdx

        mov [save_rdi],rdi
        mov [save_rsi],rsi
        mov [save_rbx],rbx
        mov [save_rbp],rbp
M
Mark Adler 已提交
181 182 183
IFDEF INFOZIP
        mov r8d,ecx
ELSE
M
Mark Adler 已提交
184
        mov r8d,edx
M
Mark Adler 已提交
185
ENDIF
M
Mark Adler 已提交
186 187 188 189 190 191 192 193 194 195 196 197
        mov [save_r12],r12
        mov [save_r13],r13
;        mov [save_r14],r14
;        mov [save_r15],r15


;;; uInt wmask = s->w_mask;
;;; unsigned chain_length = s->max_chain_length;
;;; if (s->prev_length >= s->good_match) {
;;;     chain_length >>= 2;
;;; }

M
Mark Adler 已提交
198 199 200 201
        mov edi, prev_length
        mov esi, good_match
        mov eax, WMask
        mov ebx, max_chain_length
M
Mark Adler 已提交
202 203 204 205 206 207 208 209 210 211 212 213 214 215
        cmp edi, esi
        jl  LastMatchGood
        shr ebx, 2
LastMatchGood:

;;; chainlen is decremented once beforehand so that the function can
;;; use the sign flag instead of the zero flag for the exit test.
;;; It is then shifted into the high word, to make room for the wmask
;;; value, which it will always accompany.

        dec ebx
        shl ebx, 16
        or  ebx, eax

M
Mark Adler 已提交
216
;;; on zlib only
M
Mark Adler 已提交
217 218
;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;

M
Mark Adler 已提交
219 220 221 222 223 224 225
IFDEF INFOZIP
        mov [chainlenwmask], ebx
; on infozip nice_match = [nice_match]
ELSE
        mov eax, nice_match
        mov [chainlenwmask], ebx
        mov r10d, Lookahead
M
Mark Adler 已提交
226 227 228
        cmp r10d, eax
        cmovnl r10d, eax
        mov [nicematch],r10d
M
Mark Adler 已提交
229
ENDIF
M
Mark Adler 已提交
230 231

;;; register Bytef *scan = s->window + s->strstart;
M
Mark Adler 已提交
232 233
        mov r10, window_ad
        mov ebp, strstart
M
Mark Adler 已提交
234 235 236 237 238 239 240 241 242 243 244
        lea r13, [r10 + rbp]

;;; Determine how many bytes the scan ptr is off from being
;;; dword-aligned.

         mov r9,r13
         neg r13
         and r13,3

;;; IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
;;;     s->strstart - (IPos)MAX_DIST(s) : NIL;
M
Mark Adler 已提交
245 246 247 248
IFDEF INFOZIP
        mov eax,07efah ; MAX_DIST = (WSIZE-MIN_LOOKAHEAD) (0x8000-(3+8+1))
ELSE
        mov eax, window_size
M
Mark Adler 已提交
249
        sub eax, MIN_LOOKAHEAD
M
Mark Adler 已提交
250
ENDIF
M
Mark Adler 已提交
251 252 253
        xor edi,edi
        sub ebp, eax

M
Mark Adler 已提交
254
        mov r11d, prev_length
M
Mark Adler 已提交
255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

        cmovng ebp,edi

;;; int best_len = s->prev_length;


;;; Store the sum of s->window + best_len in esi locally, and in esi.

       lea  rsi,[r10+r11]

;;; register ush scan_start = *(ushf*)scan;
;;; register ush scan_end   = *(ushf*)(scan+best_len-1);
;;; Posf *prev = s->prev;

        movzx r12d,word ptr [r9]
M
Mark Adler 已提交
270 271 272
        movzx ebx, word ptr [r9 + r11 - 1]

        mov rdi, prev_ad
M
Mark Adler 已提交
273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375

;;; Jump into the main loop.

        mov edx, [chainlenwmask]

        cmp bx,word ptr [rsi + r8 - 1]
        jz  LookupLoopIsZero

LookupLoop1:
        and r8d, edx

        movzx   r8d, word ptr [rdi + r8*2]
        cmp r8d, ebp
        jbe LeaveNow
        sub edx, 00010000h
        js  LeaveNow

LoopEntry1:
        cmp bx,word ptr [rsi + r8 - 1]
        jz  LookupLoopIsZero

LookupLoop2:
        and r8d, edx

        movzx   r8d, word ptr [rdi + r8*2]
        cmp r8d, ebp
        jbe LeaveNow
        sub edx, 00010000h
        js  LeaveNow

LoopEntry2:
        cmp bx,word ptr [rsi + r8 - 1]
        jz  LookupLoopIsZero

LookupLoop4:
        and r8d, edx

        movzx   r8d, word ptr [rdi + r8*2]
        cmp r8d, ebp
        jbe LeaveNow
        sub edx, 00010000h
        js  LeaveNow

LoopEntry4:

        cmp bx,word ptr [rsi + r8 - 1]
        jnz LookupLoop1
        jmp LookupLoopIsZero


;;; do {
;;;     match = s->window + cur_match;
;;;     if (*(ushf*)(match+best_len-1) != scan_end ||
;;;         *(ushf*)match != scan_start) continue;
;;;     [...]
;;; } while ((cur_match = prev[cur_match & wmask]) > limit
;;;          && --chain_length != 0);
;;;
;;; Here is the inner loop of the function. The function will spend the
;;; majority of its time in this loop, and majority of that time will
;;; be spent in the first ten instructions.
;;;
;;; Within this loop:
;;; ebx = scanend
;;; r8d = curmatch
;;; edx = chainlenwmask - i.e., ((chainlen << 16) | wmask)
;;; esi = windowbestlen - i.e., (window + bestlen)
;;; edi = prev
;;; ebp = limit

LookupLoop:
        and r8d, edx

        movzx   r8d, word ptr [rdi + r8*2]
        cmp r8d, ebp
        jbe LeaveNow
        sub edx, 00010000h
        js  LeaveNow

LoopEntry:

        cmp bx,word ptr [rsi + r8 - 1]
        jnz LookupLoop1
LookupLoopIsZero:
        cmp     r12w, word ptr [r10 + r8]
        jnz LookupLoop1


;;; Store the current value of chainlen.
        mov [chainlenwmask], edx

;;; Point edi to the string under scrutiny, and esi to the string we
;;; are hoping to match it up with. In actuality, esi and edi are
;;; both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and edx is
;;; initialized to -(MAX_MATCH_8 - scanalign).

        lea rsi,[r8+r10]
        mov rdx, 0fffffffffffffef8h; -(MAX_MATCH_8)
        lea rsi, [rsi + r13 + 0108h] ;MAX_MATCH_8]
        lea rdi, [r9 + r13 + 0108h] ;MAX_MATCH_8]

        prefetcht1 [rsi+rdx]
        prefetcht1 [rdi+rdx]
M
Mark Adler 已提交
376 377


M
Mark Adler 已提交
378
;;; Test the strings for equality, 8 bytes at a time. At the end,
M
Mark Adler 已提交
379
;;; adjust rdx so that it is offset to the exact byte that mismatched.
M
Mark Adler 已提交
380 381 382 383 384
;;;
;;; We already know at this point that the first three bytes of the
;;; strings match each other, and they can be safely passed over before
;;; starting the compare loop. So what this code does is skip over 0-3
;;; bytes, as much as necessary in order to dword-align the edi
M
Mark Adler 已提交
385
;;; pointer. (rsi will still be misaligned three times out of four.)
M
Mark Adler 已提交
386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
;;;
;;; It should be confessed that this loop usually does not represent
;;; much of the total running time. Replacing it with a more
;;; straightforward "rep cmpsb" would not drastically degrade
;;; performance.


LoopCmps:
        mov rax, [rsi + rdx]
        xor rax, [rdi + rdx]
        jnz LeaveLoopCmps

        mov rax, [rsi + rdx + 8]
        xor rax, [rdi + rdx + 8]
        jnz LeaveLoopCmps8


        mov rax, [rsi + rdx + 8+8]
        xor rax, [rdi + rdx + 8+8]
        jnz LeaveLoopCmps16

        add rdx,8+8+8

M
Mark Adler 已提交
409 410
        jnz short LoopCmps
        jmp short LenMaximum
M
Mark Adler 已提交
411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448
LeaveLoopCmps16: add rdx,8
LeaveLoopCmps8: add rdx,8
LeaveLoopCmps:

        test    eax, 0000FFFFh
        jnz LenLower

        test eax,0ffffffffh

        jnz LenLower32

        add rdx,4
        shr rax,32
        or ax,ax
        jnz LenLower

LenLower32:
        shr eax,16
        add rdx,2
LenLower:   sub al, 1
        adc rdx, 0
;;; Calculate the length of the match. If it is longer than MAX_MATCH,
;;; then automatically accept it as the best possible match and leave.

        lea rax, [rdi + rdx]
        sub rax, r9
        cmp eax, MAX_MATCH
        jge LenMaximum

;;; If the length of the match is not longer than the best match we
;;; have so far, then forget it and return to the lookup loop.
;///////////////////////////////////

        cmp eax, r11d
        jg  LongerMatch

        lea rsi,[r10+r11]

M
Mark Adler 已提交
449
        mov rdi, prev_ad
M
Mark Adler 已提交
450 451 452 453 454 455 456 457 458 459
        mov edx, [chainlenwmask]
        jmp LookupLoop

;;;         s->match_start = cur_match;
;;;         best_len = len;
;;;         if (len >= nice_match) break;
;;;         scan_end = *(ushf*)(scan+best_len-1);

LongerMatch:
        mov r11d, eax
M
Mark Adler 已提交
460
        mov match_start, r8d
M
Mark Adler 已提交
461 462 463 464 465 466
        cmp eax, [nicematch]
        jge LeaveNow

        lea rsi,[r10+rax]

        movzx   ebx, word ptr [r9 + rax - 1]
M
Mark Adler 已提交
467
        mov rdi, prev_ad
M
Mark Adler 已提交
468 469 470 471 472 473 474
        mov edx, [chainlenwmask]
        jmp LookupLoop

;;; Accept the current string, with the maximum possible length.

LenMaximum:
        mov r11d,MAX_MATCH
M
Mark Adler 已提交
475
        mov match_start, r8d
M
Mark Adler 已提交
476 477 478 479 480

;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len;
;;; return s->lookahead;

LeaveNow:
M
Mark Adler 已提交
481 482 483 484
IFDEF INFOZIP
        mov eax,r11d
ELSE
        mov eax, Lookahead
M
Mark Adler 已提交
485 486
        cmp r11d, eax
        cmovng eax, r11d
M
Mark Adler 已提交
487
ENDIF
M
Mark Adler 已提交
488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503

;;; Restore the stack and return from whence we came.


        mov rsi,[save_rsi]
        mov rdi,[save_rdi]
        mov rbx,[save_rbx]
        mov rbp,[save_rbp]
        mov r12,[save_r12]
        mov r13,[save_r13]
;        mov r14,[save_r14]
;        mov r15,[save_r15]


        ret 0
; please don't remove this string !
M
Mark Adler 已提交
504 505
; Your can freely use gvmat64 in any free or commercial app
; but it is far better don't remove the string in the binary!
M
Mark Adler 已提交
506 507 508 509 510 511 512 513 514
    db     0dh,0ah,"asm686 with masm, optimised assembly code from Brian Raiter, written 1998, converted to amd 64 by Gilles Vollant 2005",0dh,0ah,0
longest_match   ENDP

match_init PROC
  ret 0
match_init ENDP


END