From 6b8233bfe00e79134cb1b84fc49d4f750a797f79 Mon Sep 17 00:00:00 2001 From: Mark Adler Date: Fri, 9 Sep 2011 23:24:43 -0700 Subject: zlib 1.2.2.3 --- contrib/masmx64/gvmat64.asm | 464 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 464 insertions(+) create mode 100644 contrib/masmx64/gvmat64.asm (limited to 'contrib/masmx64/gvmat64.asm') diff --git a/contrib/masmx64/gvmat64.asm b/contrib/masmx64/gvmat64.asm new file mode 100644 index 0000000..cee2145 --- /dev/null +++ b/contrib/masmx64/gvmat64.asm @@ -0,0 +1,464 @@ +;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. +; File written by Gilles Vollant, by modifiying the longest_match +; from Jean-loup Gailly in deflate.c +; and modifying asm686 with masm, optimised assembly code from Brian Raiter, +; written 1998 +; http://www.zlib.net +; http://www.winimage.com/zLibDll +; http://www.muppetlabs.com/~breadbox/software/assembly.html +; +; to compile this file, I use option +; ml64.exe /Flgvmat64 /c /Zi gvmat64.asm +; with Microsoft Macro Assembler (x64) for AMD64 +; +; ml64.exe is given with Visual Studio 2005, Windows 2003 server DDK +; +; (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) +; +; Be carrefull to adapt zlib1222add below to your version of zLib + + +;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 + nicematch equ rsp + 16 - LocalVarsSize ; a good enough match size -> r14 + +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"). + +zlib1222add equ 8 + +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 + + +; 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 +; rax, rcx, rdx, r8, r-9, r10, and r11, which are scratch. + + + +;;; 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 + mov r8d,edx + 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; +;;; } + + mov edi, [rcx + dsPrevLen] + mov esi, [rcx + dsGoodMatch] + mov eax, [rcx + dsWMask] + mov ebx, [rcx + dsMaxChainLen] + 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 + mov [chainlenwmask], ebx + +;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead; + + mov eax, [rcx + dsNiceMatch] + mov r10d, [rcx + dsLookahead] + cmp r10d, eax + cmovnl r10d, eax + mov [nicematch],r10d +LookaheadLess: + +;;; register Bytef *scan = s->window + s->strstart; + + mov r10, [rcx + dsWindow] + mov ebp, [rcx + dsStrStart] + 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; + + mov eax, [rcx + dsWSize] + sub eax, MIN_LOOKAHEAD + xor edi,edi + sub ebp, eax + + mov r11d, [rcx + dsPrevLen] + + 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] + movzx ebx, word ptr [r9 + r11 - 1] + mov rdi, [rcx + dsPrev] + +;;; 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] +;;; Test the strings for equality, 8 bytes at a time. At the end, +;;; adjust edx so that it is offset to the exact byte that mismatched. +;;; +;;; 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 +;;; pointer. (esi will still be misaligned three times out of four.) +;;; +;;; 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 eax, [rsi + rdx] +; xor eax, [rdi + rdx] +; jnz LeaveLoopCmps +; mov eax, [rsi + rdx + 4] +; xor eax, [rdi + rdx + 4] +; jnz LeaveLoopCmps4 +; add rdx, 8 +; jnz LoopCmps +; jmp LenMaximum +;LeaveLoopCmps4: add rdx, 4 +;LeaveLoopCmps: test eax, 0000FFFFh +; jnz LenLower +; add rdx, 2 +; shr eax, 16 +;LenLower: sub al, 1 +; adc rdx, 0 + + +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 + + jmp short LoopCmps +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] + + mov rdi, [rcx + dsPrev] + 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 + mov [rcx + dsMatchStart], r8d + cmp eax, [nicematch] + jge LeaveNow + + lea rsi,[r10+rax] + + movzx ebx, word ptr [r9 + rax - 1] + mov rdi, [rcx + dsPrev] + mov edx, [chainlenwmask] + jmp LookupLoop + +;;; Accept the current string, with the maximum possible length. + +LenMaximum: + mov r11d,MAX_MATCH + mov [rcx + dsMatchStart], r8d + +;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len; +;;; return s->lookahead; + +LeaveNow: + mov eax, [rcx + dsLookahead] + cmp r11d, eax + cmovng eax, r11d + + +;;; 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 ! +; Your can freely use gvmat32 in any free or commercial app if you don't remove the string in the binary! + 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 -- cgit v1.2.3-55-g6feb