+; match686.asm -- Asm portion of the optimized longest_match for 32 bits x86\r
+; Copyright (C) 1995-1996 Jean-loup Gailly, Brian Raiter and Gilles Vollant.\r
+; File written by Gilles Vollant, by converting match686.S from Brian Raiter\r
+; for MASM. This is as assembly version of longest_match\r
+; from Jean-loup Gailly in deflate.c\r
+;\r
+; http://www.zlib.net\r
+; http://www.winimage.com/zLibDll\r
+; http://www.muppetlabs.com/~breadbox/software/assembly.html\r
+;\r
+; For Visual C++ 4.x and higher and ML 6.x and higher\r
+; ml.exe is distributed in\r
+; http://www.microsoft.com/downloads/details.aspx?FamilyID=7a1c9da0-0510-44a2-b042-7ef370530c64\r
+;\r
+; this file contain two implementation of longest_match\r
+;\r
+; this longest_match was written by Brian raiter (1998), optimized for Pentium Pro\r
+; (and the faster known version of match_init on modern Core 2 Duo and AMD Phenom)\r
+;\r
+; for using an assembly version of longest_match, you need define ASMV in project\r
+;\r
+; compile the asm file running\r
+; ml /coff /Zi /c /Flmatch686.lst match686.asm\r
+; and do not include match686.obj in your project\r
+;\r
+; note: contrib of zLib 1.2.3 and earlier contained both a deprecated version for\r
+; Pentium (prior Pentium Pro) and this version for Pentium Pro and modern processor\r
+; with autoselect (with cpu detection code)\r
+; if you want support the old pentium optimization, you can still use these version\r
+;\r
+; this file is not optimized for old pentium, but it compatible with all x86 32 bits\r
+; processor (starting 80386)\r
+;\r
+;\r
+; see below : zlib1222add must be adjuster if you use a zlib version < 1.2.2.2\r
+\r
+;uInt longest_match(s, cur_match)\r
+; deflate_state *s;\r
+; IPos cur_match; /* current match */\r
+\r
+ NbStack equ 76\r
+ cur_match equ dword ptr[esp+NbStack-0]\r
+ str_s equ dword ptr[esp+NbStack-4]\r
+; 5 dword on top (ret,ebp,esi,edi,ebx)\r
+ adrret equ dword ptr[esp+NbStack-8]\r
+ pushebp equ dword ptr[esp+NbStack-12]\r
+ pushedi equ dword ptr[esp+NbStack-16]\r
+ pushesi equ dword ptr[esp+NbStack-20]\r
+ pushebx equ dword ptr[esp+NbStack-24]\r
+\r
+ chain_length equ dword ptr [esp+NbStack-28]\r
+ limit equ dword ptr [esp+NbStack-32]\r
+ best_len equ dword ptr [esp+NbStack-36]\r
+ window equ dword ptr [esp+NbStack-40]\r
+ prev equ dword ptr [esp+NbStack-44]\r
+ scan_start equ word ptr [esp+NbStack-48]\r
+ wmask equ dword ptr [esp+NbStack-52]\r
+ match_start_ptr equ dword ptr [esp+NbStack-56]\r
+ nice_match equ dword ptr [esp+NbStack-60]\r
+ scan equ dword ptr [esp+NbStack-64]\r
+\r
+ windowlen equ dword ptr [esp+NbStack-68]\r
+ match_start equ dword ptr [esp+NbStack-72]\r
+ strend equ dword ptr [esp+NbStack-76]\r
+ NbStackAdd equ (NbStack-24)\r
+\r
+ .386p\r
+\r
+ name gvmatch\r
+ .MODEL FLAT\r
+\r
+\r
+\r
+; all the +zlib1222add offsets are due to the addition of fields\r
+; in zlib in the deflate_state structure since the asm code was first written\r
+; (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)").\r
+; (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0").\r
+; if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8").\r
+\r
+ zlib1222add equ 8\r
+\r
+; Note : these value are good with a 8 bytes boundary pack structure\r
+ dep_chain_length equ 74h+zlib1222add\r
+ dep_window equ 30h+zlib1222add\r
+ dep_strstart equ 64h+zlib1222add\r
+ dep_prev_length equ 70h+zlib1222add\r
+ dep_nice_match equ 88h+zlib1222add\r
+ dep_w_size equ 24h+zlib1222add\r
+ dep_prev equ 38h+zlib1222add\r
+ dep_w_mask equ 2ch+zlib1222add\r
+ dep_good_match equ 84h+zlib1222add\r
+ dep_match_start equ 68h+zlib1222add\r
+ dep_lookahead equ 6ch+zlib1222add\r
+\r
+\r
+_TEXT segment\r
+\r
+IFDEF NOUNDERLINE\r
+ public longest_match\r
+ public match_init\r
+ELSE\r
+ public _longest_match\r
+ public _match_init\r
+ENDIF\r
+\r
+ MAX_MATCH equ 258\r
+ MIN_MATCH equ 3\r
+ MIN_LOOKAHEAD equ (MAX_MATCH+MIN_MATCH+1)\r
+\r
+\r
+\r
+MAX_MATCH equ 258\r
+MIN_MATCH equ 3\r
+MIN_LOOKAHEAD equ (MAX_MATCH + MIN_MATCH + 1)\r
+MAX_MATCH_8_ equ ((MAX_MATCH + 7) AND 0FFF0h)\r
+\r
+\r
+;;; stack frame offsets\r
+\r
+chainlenwmask equ esp + 0 ; high word: current chain len\r
+ ; low word: s->wmask\r
+window equ esp + 4 ; local copy of s->window\r
+windowbestlen equ esp + 8 ; s->window + bestlen\r
+scanstart equ esp + 16 ; first two bytes of string\r
+scanend equ esp + 12 ; last two bytes of string\r
+scanalign equ esp + 20 ; dword-misalignment of string\r
+nicematch equ esp + 24 ; a good enough match size\r
+bestlen equ esp + 28 ; size of best match so far\r
+scan equ esp + 32 ; ptr to string wanting match\r
+\r
+LocalVarsSize equ 36\r
+; saved ebx byte esp + 36\r
+; saved edi byte esp + 40\r
+; saved esi byte esp + 44\r
+; saved ebp byte esp + 48\r
+; return address byte esp + 52\r
+deflatestate equ esp + 56 ; the function arguments\r
+curmatch equ esp + 60\r
+\r
+;;; Offsets for fields in the deflate_state structure. These numbers\r
+;;; are calculated from the definition of deflate_state, with the\r
+;;; assumption that the compiler will dword-align the fields. (Thus,\r
+;;; changing the definition of deflate_state could easily cause this\r
+;;; program to crash horribly, without so much as a warning at\r
+;;; compile time. Sigh.)\r
+\r
+dsWSize equ 36+zlib1222add\r
+dsWMask equ 44+zlib1222add\r
+dsWindow equ 48+zlib1222add\r
+dsPrev equ 56+zlib1222add\r
+dsMatchLen equ 88+zlib1222add\r
+dsPrevMatch equ 92+zlib1222add\r
+dsStrStart equ 100+zlib1222add\r
+dsMatchStart equ 104+zlib1222add\r
+dsLookahead equ 108+zlib1222add\r
+dsPrevLen equ 112+zlib1222add\r
+dsMaxChainLen equ 116+zlib1222add\r
+dsGoodMatch equ 132+zlib1222add\r
+dsNiceMatch equ 136+zlib1222add\r
+\r
+\r
+;;; match686.asm -- Pentium-Pro-optimized version of longest_match()\r
+;;; Written for zlib 1.1.2\r
+;;; Copyright (C) 1998 Brian Raiter <breadbox@muppetlabs.com>\r
+;;; You can look at http://www.muppetlabs.com/~breadbox/software/assembly.html\r
+;;;\r
+;;\r
+;; This software is provided 'as-is', without any express or implied\r
+;; warranty. In no event will the authors be held liable for any damages\r
+;; arising from the use of this software.\r
+;;\r
+;; Permission is granted to anyone to use this software for any purpose,\r
+;; including commercial applications, and to alter it and redistribute it\r
+;; freely, subject to the following restrictions:\r
+;;\r
+;; 1. The origin of this software must not be misrepresented; you must not\r
+;; claim that you wrote the original software. If you use this software\r
+;; in a product, an acknowledgment in the product documentation would be\r
+;; appreciated but is not required.\r
+;; 2. Altered source versions must be plainly marked as such, and must not be\r
+;; misrepresented as being the original software\r
+;; 3. This notice may not be removed or altered from any source distribution.\r
+;;\r
+\r
+;GLOBAL _longest_match, _match_init\r
+\r
+\r
+;SECTION .text\r
+\r
+;;; uInt longest_match(deflate_state *deflatestate, IPos curmatch)\r
+\r
+;_longest_match:\r
+ IFDEF NOUNDERLINE\r
+ longest_match proc near\r
+ ELSE\r
+ _longest_match proc near\r
+ ENDIF\r
+.FPO (9, 4, 0, 0, 1, 0)\r
+\r
+;;; Save registers that the compiler may be using, and adjust esp to\r
+;;; make room for our stack frame.\r
+\r
+ push ebp\r
+ push edi\r
+ push esi\r
+ push ebx\r
+ sub esp, LocalVarsSize\r
+\r
+;;; Retrieve the function arguments. ecx will hold cur_match\r
+;;; throughout the entire function. edx will hold the pointer to the\r
+;;; deflate_state structure during the function's setup (before\r
+;;; entering the main loop.\r
+\r
+ mov edx, [deflatestate]\r
+ mov ecx, [curmatch]\r
+\r
+;;; uInt wmask = s->w_mask;\r
+;;; unsigned chain_length = s->max_chain_length;\r
+;;; if (s->prev_length >= s->good_match) {\r
+;;; chain_length >>= 2;\r
+;;; }\r
+\r
+ mov eax, [edx + dsPrevLen]\r
+ mov ebx, [edx + dsGoodMatch]\r
+ cmp eax, ebx\r
+ mov eax, [edx + dsWMask]\r
+ mov ebx, [edx + dsMaxChainLen]\r
+ jl LastMatchGood\r
+ shr ebx, 2\r
+LastMatchGood:\r
+\r
+;;; chainlen is decremented once beforehand so that the function can\r
+;;; use the sign flag instead of the zero flag for the exit test.\r
+;;; It is then shifted into the high word, to make room for the wmask\r
+;;; value, which it will always accompany.\r
+\r
+ dec ebx\r
+ shl ebx, 16\r
+ or ebx, eax\r
+ mov [chainlenwmask], ebx\r
+\r
+;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;\r
+\r
+ mov eax, [edx + dsNiceMatch]\r
+ mov ebx, [edx + dsLookahead]\r
+ cmp ebx, eax\r
+ jl LookaheadLess\r
+ mov ebx, eax\r
+LookaheadLess: mov [nicematch], ebx\r
+\r
+;;; register Bytef *scan = s->window + s->strstart;\r
+\r
+ mov esi, [edx + dsWindow]\r
+ mov [window], esi\r
+ mov ebp, [edx + dsStrStart]\r
+ lea edi, [esi + ebp]\r
+ mov [scan], edi\r
+\r
+;;; Determine how many bytes the scan ptr is off from being\r
+;;; dword-aligned.\r
+\r
+ mov eax, edi\r
+ neg eax\r
+ and eax, 3\r
+ mov [scanalign], eax\r
+\r
+;;; IPos limit = s->strstart > (IPos)MAX_DIST(s) ?\r
+;;; s->strstart - (IPos)MAX_DIST(s) : NIL;\r
+\r
+ mov eax, [edx + dsWSize]\r
+ sub eax, MIN_LOOKAHEAD\r
+ sub ebp, eax\r
+ jg LimitPositive\r
+ xor ebp, ebp\r
+LimitPositive:\r
+\r
+;;; int best_len = s->prev_length;\r
+\r
+ mov eax, [edx + dsPrevLen]\r
+ mov [bestlen], eax\r
+\r
+;;; Store the sum of s->window + best_len in esi locally, and in esi.\r
+\r
+ add esi, eax\r
+ mov [windowbestlen], esi\r
+\r
+;;; register ush scan_start = *(ushf*)scan;\r
+;;; register ush scan_end = *(ushf*)(scan+best_len-1);\r
+;;; Posf *prev = s->prev;\r
+\r
+ movzx ebx, word ptr [edi]\r
+ mov [scanstart], ebx\r
+ movzx ebx, word ptr [edi + eax - 1]\r
+ mov [scanend], ebx\r
+ mov edi, [edx + dsPrev]\r
+\r
+;;; Jump into the main loop.\r
+\r
+ mov edx, [chainlenwmask]\r
+ jmp short LoopEntry\r
+\r
+align 4\r
+\r
+;;; do {\r
+;;; match = s->window + cur_match;\r
+;;; if (*(ushf*)(match+best_len-1) != scan_end ||\r
+;;; *(ushf*)match != scan_start) continue;\r
+;;; [...]\r
+;;; } while ((cur_match = prev[cur_match & wmask]) > limit\r
+;;; && --chain_length != 0);\r
+;;;\r
+;;; Here is the inner loop of the function. The function will spend the\r
+;;; majority of its time in this loop, and majority of that time will\r
+;;; be spent in the first ten instructions.\r
+;;;\r
+;;; Within this loop:\r
+;;; ebx = scanend\r
+;;; ecx = curmatch\r
+;;; edx = chainlenwmask - i.e., ((chainlen << 16) | wmask)\r
+;;; esi = windowbestlen - i.e., (window + bestlen)\r
+;;; edi = prev\r
+;;; ebp = limit\r
+\r
+LookupLoop:\r
+ and ecx, edx\r
+ movzx ecx, word ptr [edi + ecx*2]\r
+ cmp ecx, ebp\r
+ jbe LeaveNow\r
+ sub edx, 00010000h\r
+ js LeaveNow\r
+LoopEntry: movzx eax, word ptr [esi + ecx - 1]\r
+ cmp eax, ebx\r
+ jnz LookupLoop\r
+ mov eax, [window]\r
+ movzx eax, word ptr [eax + ecx]\r
+ cmp eax, [scanstart]\r
+ jnz LookupLoop\r
+\r
+;;; Store the current value of chainlen.\r
+\r
+ mov [chainlenwmask], edx\r
+\r
+;;; Point edi to the string under scrutiny, and esi to the string we\r
+;;; are hoping to match it up with. In actuality, esi and edi are\r
+;;; both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and edx is\r
+;;; initialized to -(MAX_MATCH_8 - scanalign).\r
+\r
+ mov esi, [window]\r
+ mov edi, [scan]\r
+ add esi, ecx\r
+ mov eax, [scanalign]\r
+ mov edx, 0fffffef8h; -(MAX_MATCH_8)\r
+ lea edi, [edi + eax + 0108h] ;MAX_MATCH_8]\r
+ lea esi, [esi + eax + 0108h] ;MAX_MATCH_8]\r
+\r
+;;; Test the strings for equality, 8 bytes at a time. At the end,\r
+;;; adjust edx so that it is offset to the exact byte that mismatched.\r
+;;;\r
+;;; We already know at this point that the first three bytes of the\r
+;;; strings match each other, and they can be safely passed over before\r
+;;; starting the compare loop. So what this code does is skip over 0-3\r
+;;; bytes, as much as necessary in order to dword-align the edi\r
+;;; pointer. (esi will still be misaligned three times out of four.)\r
+;;;\r
+;;; It should be confessed that this loop usually does not represent\r
+;;; much of the total running time. Replacing it with a more\r
+;;; straightforward "rep cmpsb" would not drastically degrade\r
+;;; performance.\r
+\r
+LoopCmps:\r
+ mov eax, [esi + edx]\r
+ xor eax, [edi + edx]\r
+ jnz LeaveLoopCmps\r
+ mov eax, [esi + edx + 4]\r
+ xor eax, [edi + edx + 4]\r
+ jnz LeaveLoopCmps4\r
+ add edx, 8\r
+ jnz LoopCmps\r
+ jmp short LenMaximum\r
+LeaveLoopCmps4: add edx, 4\r
+LeaveLoopCmps: test eax, 0000FFFFh\r
+ jnz LenLower\r
+ add edx, 2\r
+ shr eax, 16\r
+LenLower: sub al, 1\r
+ adc edx, 0\r
+\r
+;;; Calculate the length of the match. If it is longer than MAX_MATCH,\r
+;;; then automatically accept it as the best possible match and leave.\r
+\r
+ lea eax, [edi + edx]\r
+ mov edi, [scan]\r
+ sub eax, edi\r
+ cmp eax, MAX_MATCH\r
+ jge LenMaximum\r
+\r
+;;; If the length of the match is not longer than the best match we\r
+;;; have so far, then forget it and return to the lookup loop.\r
+\r
+ mov edx, [deflatestate]\r
+ mov ebx, [bestlen]\r
+ cmp eax, ebx\r
+ jg LongerMatch\r
+ mov esi, [windowbestlen]\r
+ mov edi, [edx + dsPrev]\r
+ mov ebx, [scanend]\r
+ mov edx, [chainlenwmask]\r
+ jmp LookupLoop\r
+\r
+;;; s->match_start = cur_match;\r
+;;; best_len = len;\r
+;;; if (len >= nice_match) break;\r
+;;; scan_end = *(ushf*)(scan+best_len-1);\r
+\r
+LongerMatch: mov ebx, [nicematch]\r
+ mov [bestlen], eax\r
+ mov [edx + dsMatchStart], ecx\r
+ cmp eax, ebx\r
+ jge LeaveNow\r
+ mov esi, [window]\r
+ add esi, eax\r
+ mov [windowbestlen], esi\r
+ movzx ebx, word ptr [edi + eax - 1]\r
+ mov edi, [edx + dsPrev]\r
+ mov [scanend], ebx\r
+ mov edx, [chainlenwmask]\r
+ jmp LookupLoop\r
+\r
+;;; Accept the current string, with the maximum possible length.\r
+\r
+LenMaximum: mov edx, [deflatestate]\r
+ mov dword ptr [bestlen], MAX_MATCH\r
+ mov [edx + dsMatchStart], ecx\r
+\r
+;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len;\r
+;;; return s->lookahead;\r
+\r
+LeaveNow:\r
+ mov edx, [deflatestate]\r
+ mov ebx, [bestlen]\r
+ mov eax, [edx + dsLookahead]\r
+ cmp ebx, eax\r
+ jg LookaheadRet\r
+ mov eax, ebx\r
+LookaheadRet:\r
+\r
+;;; Restore the stack and return from whence we came.\r
+\r
+ add esp, LocalVarsSize\r
+ pop ebx\r
+ pop esi\r
+ pop edi\r
+ pop ebp\r
+\r
+ ret\r
+; please don't remove this string !\r
+; Your can freely use match686 in any free or commercial app if you don't remove the string in the binary!\r
+ db 0dh,0ah,"asm686 with masm, optimised assembly code from Brian Raiter, written 1998",0dh,0ah\r
+\r
+\r
+ IFDEF NOUNDERLINE\r
+ longest_match endp\r
+ ELSE\r
+ _longest_match endp\r
+ ENDIF\r
+\r
+ IFDEF NOUNDERLINE\r
+ match_init proc near\r
+ ret\r
+ match_init endp\r
+ ELSE\r
+ _match_init proc near\r
+ ret\r
+ _match_init endp\r
+ ENDIF\r
+\r
+\r
+_TEXT ends\r
+end\r