From 07162fa301835387e4da4278326ec1ff45d1dc3e Mon Sep 17 00:00:00 2001 From: Aki Date: Wed, 29 Sep 2021 21:52:57 +0200 Subject: Squashed 'zlib/' content from commit cacf7f1 git-subtree-dir: zlib git-subtree-split: cacf7f1d4e3d44d871b605da3b647f07d718623f --- contrib/masmx64/bld_ml64.bat | 2 + contrib/masmx64/gvmat64.asm | 553 ++++++++++++++++++++++++++++++++++++++++++ contrib/masmx64/inffas8664.c | 186 ++++++++++++++ contrib/masmx64/inffasx64.asm | 396 ++++++++++++++++++++++++++++++ contrib/masmx64/readme.txt | 31 +++ 5 files changed, 1168 insertions(+) create mode 100644 contrib/masmx64/bld_ml64.bat create mode 100644 contrib/masmx64/gvmat64.asm create mode 100644 contrib/masmx64/inffas8664.c create mode 100644 contrib/masmx64/inffasx64.asm create mode 100644 contrib/masmx64/readme.txt (limited to 'contrib/masmx64') diff --git a/contrib/masmx64/bld_ml64.bat b/contrib/masmx64/bld_ml64.bat new file mode 100644 index 0000000..8f9343d --- /dev/null +++ b/contrib/masmx64/bld_ml64.bat @@ -0,0 +1,2 @@ +ml64.exe /Flinffasx64 /c /Zi inffasx64.asm +ml64.exe /Flgvmat64 /c /Zi gvmat64.asm diff --git a/contrib/masmx64/gvmat64.asm b/contrib/masmx64/gvmat64.asm new file mode 100644 index 0000000..9879c28 --- /dev/null +++ b/contrib/masmx64/gvmat64.asm @@ -0,0 +1,553 @@ +;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_64 +; (AMD64 on Athlon 64, Opteron, Phenom +; and Intel EM64T on Pentium 4 with EM64T, Pentium D, Core 2 Duo, Core I5/I7) +; Copyright (C) 1995-2010 Jean-loup Gailly, Brian Raiter and Gilles Vollant. +; +; 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 +; +; This software is provided 'as-is', without any express or implied +; warranty. In no event will the authors be held liable for any damages +; arising from the use of this software. +; +; Permission is granted to anyone to use this software for any purpose, +; including commercial applications, and to alter it and redistribute it +; freely, subject to the following restrictions: +; +; 1. The origin of this software must not be misrepresented; you must not +; claim that you wrote the original software. If you use this software +; in a product, an acknowledgment in the product documentation would be +; appreciated but is not required. +; 2. Altered source versions must be plainly marked as such, and must not be +; misrepresented as being the original software +; 3. This notice may not be removed or altered from any source distribution. +; +; +; +; http://www.zlib.net +; http://www.winimage.com/zLibDll +; http://www.muppetlabs.com/~breadbox/software/assembly.html +; +; 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: +; ml64.exe /Flgvmat64 /c /Zi gvmat64.asm +; 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) +; +; This file compile with Microsoft Macro Assembler (x64) for AMD64 +; +; ml64.exe is given with Visual Studio 2005/2008/2010 and Windows WDK +; +; (you can get Windows WDK with ml64 for AMD64 from +; http://www.microsoft.com/whdc/Devtools/wdk/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 +IFDEF INFOZIP +ELSE + nicematch equ (rsp + 16 - LocalVarsSize) ; a good enough match size +ENDIF + +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 + + +; summary of register usage +; scanend ebx +; scanendw bx +; chainlenwmask edx +; curmatch rsi +; curmatchd esi +; windowbestlen r8 +; scanalign r9 +; scanalignd r9d +; window r10 +; bestlen r11 +; bestlend r11d +; scanstart r12d +; scanstartw r12w +; scan r13 +; nicematch r14d +; limit r15 +; limitd r15d +; prev rcx + +; 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"). + + +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 +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 + +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 + +; 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, r9, 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 +IFDEF INFOZIP + mov r8d,ecx +ELSE + mov r8d,edx +ENDIF + 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, prev_length + mov esi, good_match + mov eax, WMask + mov ebx, max_chain_length + 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 + +;;; on zlib only +;;; if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead; + +IFDEF INFOZIP + mov [chainlenwmask], ebx +; on infozip nice_match = [nice_match] +ELSE + mov eax, nice_match + mov [chainlenwmask], ebx + mov r10d, Lookahead + cmp r10d, eax + cmovnl r10d, eax + mov [nicematch],r10d +ENDIF + +;;; register Bytef *scan = s->window + s->strstart; + mov r10, window_ad + mov ebp, strstart + 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; +IFDEF INFOZIP + mov eax,07efah ; MAX_DIST = (WSIZE-MIN_LOOKAHEAD) (0x8000-(3+8+1)) +ELSE + mov eax, window_size + sub eax, MIN_LOOKAHEAD +ENDIF + xor edi,edi + sub ebp, eax + + mov r11d, prev_length + + 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, prev_ad + +;;; 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 rdx 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. (rsi 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 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 + + jnz short LoopCmps + jmp short LenMaximum +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, prev_ad + 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 match_start, r8d + cmp eax, [nicematch] + jge LeaveNow + + lea rsi,[r10+rax] + + movzx ebx, word ptr [r9 + rax - 1] + mov rdi, prev_ad + mov edx, [chainlenwmask] + jmp LookupLoop + +;;; Accept the current string, with the maximum possible length. + +LenMaximum: + mov r11d,MAX_MATCH + mov match_start, r8d + +;;; if ((uInt)best_len <= s->lookahead) return (uInt)best_len; +;;; return s->lookahead; + +LeaveNow: +IFDEF INFOZIP + mov eax,r11d +ELSE + mov eax, Lookahead + cmp r11d, eax + cmovng eax, r11d +ENDIF + +;;; 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 gvmat64 in any free or commercial app +; but it is far better 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 diff --git a/contrib/masmx64/inffas8664.c b/contrib/masmx64/inffas8664.c new file mode 100644 index 0000000..e8af06f --- /dev/null +++ b/contrib/masmx64/inffas8664.c @@ -0,0 +1,186 @@ +/* inffas8664.c is a hand tuned assembler version of inffast.c - fast decoding + * version for AMD64 on Windows using Microsoft C compiler + * + * Copyright (C) 1995-2003 Mark Adler + * For conditions of distribution and use, see copyright notice in zlib.h + * + * Copyright (C) 2003 Chris Anderson + * Please use the copyright conditions above. + * + * 2005 - Adaptation to Microsoft C Compiler for AMD64 by Gilles Vollant + * + * inffas8664.c call function inffas8664fnc in inffasx64.asm + * inffasx64.asm is automatically convert from AMD64 portion of inffas86.c + * + * Dec-29-2003 -- I added AMD64 inflate asm support. This version is also + * slightly quicker on x86 systems because, instead of using rep movsb to copy + * data, it uses rep movsw, which moves data in 2-byte chunks instead of single + * bytes. I've tested the AMD64 code on a Fedora Core 1 + the x86_64 updates + * from http://fedora.linux.duke.edu/fc1_x86_64 + * which is running on an Athlon 64 3000+ / Gigabyte GA-K8VT800M system with + * 1GB ram. The 64-bit version is about 4% faster than the 32-bit version, + * when decompressing mozilla-source-1.3.tar.gz. + * + * Mar-13-2003 -- Most of this is derived from inffast.S which is derived from + * the gcc -S output of zlib-1.2.0/inffast.c. Zlib-1.2.0 is in beta release at + * the moment. I have successfully compiled and tested this code with gcc2.96, + * gcc3.2, icc5.0, msvc6.0. It is very close to the speed of inffast.S + * compiled with gcc -DNO_MMX, but inffast.S is still faster on the P3 with MMX + * enabled. I will attempt to merge the MMX code into this version. Newer + * versions of this and inffast.S can be found at + * http://www.eetbeetee.com/zlib/ and http://www.charm.net/~christop/zlib/ + * + */ + +#include +#include "zutil.h" +#include "inftrees.h" +#include "inflate.h" +#include "inffast.h" + +/* Mark Adler's comments from inffast.c: */ + +/* + Decode literal, length, and distance codes and write out the resulting + literal and match bytes until either not enough input or output is + available, an end-of-block is encountered, or a data error is encountered. + When large enough input and output buffers are supplied to inflate(), for + example, a 16K input buffer and a 64K output buffer, more than 95% of the + inflate execution time is spent in this routine. + + Entry assumptions: + + state->mode == LEN + strm->avail_in >= 6 + strm->avail_out >= 258 + start >= strm->avail_out + state->bits < 8 + + On return, state->mode is one of: + + LEN -- ran out of enough output space or enough available input + TYPE -- reached end of block code, inflate() to interpret next block + BAD -- error in block data + + Notes: + + - The maximum input bits used by a length/distance pair is 15 bits for the + length code, 5 bits for the length extra, 15 bits for the distance code, + and 13 bits for the distance extra. This totals 48 bits, or six bytes. + Therefore if strm->avail_in >= 6, then there is enough input to avoid + checking for available input while decoding. + + - The maximum bytes that a single length/distance pair can output is 258 + bytes, which is the maximum length that can be coded. inflate_fast() + requires strm->avail_out >= 258 for each loop to avoid checking for + output space. + */ + + + + typedef struct inffast_ar { +/* 64 32 x86 x86_64 */ +/* ar offset register */ +/* 0 0 */ void *esp; /* esp save */ +/* 8 4 */ void *ebp; /* ebp save */ +/* 16 8 */ unsigned char FAR *in; /* esi rsi local strm->next_in */ +/* 24 12 */ unsigned char FAR *last; /* r9 while in < last */ +/* 32 16 */ unsigned char FAR *out; /* edi rdi local strm->next_out */ +/* 40 20 */ unsigned char FAR *beg; /* inflate()'s init next_out */ +/* 48 24 */ unsigned char FAR *end; /* r10 while out < end */ +/* 56 28 */ unsigned char FAR *window;/* size of window, wsize!=0 */ +/* 64 32 */ code const FAR *lcode; /* ebp rbp local strm->lencode */ +/* 72 36 */ code const FAR *dcode; /* r11 local strm->distcode */ +/* 80 40 */ size_t /*unsigned long */hold; /* edx rdx local strm->hold */ +/* 88 44 */ unsigned bits; /* ebx rbx local strm->bits */ +/* 92 48 */ unsigned wsize; /* window size */ +/* 96 52 */ unsigned write; /* window write index */ +/*100 56 */ unsigned lmask; /* r12 mask for lcode */ +/*104 60 */ unsigned dmask; /* r13 mask for dcode */ +/*108 64 */ unsigned len; /* r14 match length */ +/*112 68 */ unsigned dist; /* r15 match distance */ +/*116 72 */ unsigned status; /* set when state chng*/ + } type_ar; +#ifdef ASMINF + +void inflate_fast(strm, start) +z_streamp strm; +unsigned start; /* inflate()'s starting value for strm->avail_out */ +{ + struct inflate_state FAR *state; + type_ar ar; + void inffas8664fnc(struct inffast_ar * par); + + + +#if (defined( __GNUC__ ) && defined( __amd64__ ) && ! defined( __i386 )) || (defined(_MSC_VER) && defined(_M_AMD64)) +#define PAD_AVAIL_IN 6 +#define PAD_AVAIL_OUT 258 +#else +#define PAD_AVAIL_IN 5 +#define PAD_AVAIL_OUT 257 +#endif + + /* copy state to local variables */ + state = (struct inflate_state FAR *)strm->state; + + ar.in = strm->next_in; + ar.last = ar.in + (strm->avail_in - PAD_AVAIL_IN); + ar.out = strm->next_out; + ar.beg = ar.out - (start - strm->avail_out); + ar.end = ar.out + (strm->avail_out - PAD_AVAIL_OUT); + ar.wsize = state->wsize; + ar.write = state->wnext; + ar.window = state->window; + ar.hold = state->hold; + ar.bits = state->bits; + ar.lcode = state->lencode; + ar.dcode = state->distcode; + ar.lmask = (1U << state->lenbits) - 1; + ar.dmask = (1U << state->distbits) - 1; + + /* decode literals and length/distances until end-of-block or not enough + input data or output space */ + + /* align in on 1/2 hold size boundary */ + while (((size_t)(void *)ar.in & (sizeof(ar.hold) / 2 - 1)) != 0) { + ar.hold += (unsigned long)*ar.in++ << ar.bits; + ar.bits += 8; + } + + inffas8664fnc(&ar); + + if (ar.status > 1) { + if (ar.status == 2) + strm->msg = "invalid literal/length code"; + else if (ar.status == 3) + strm->msg = "invalid distance code"; + else + strm->msg = "invalid distance too far back"; + state->mode = BAD; + } + else if ( ar.status == 1 ) { + state->mode = TYPE; + } + + /* return unused bytes (on entry, bits < 8, so in won't go too far back) */ + ar.len = ar.bits >> 3; + ar.in -= ar.len; + ar.bits -= ar.len << 3; + ar.hold &= (1U << ar.bits) - 1; + + /* update state and return */ + strm->next_in = ar.in; + strm->next_out = ar.out; + strm->avail_in = (unsigned)(ar.in < ar.last ? + PAD_AVAIL_IN + (ar.last - ar.in) : + PAD_AVAIL_IN - (ar.in - ar.last)); + strm->avail_out = (unsigned)(ar.out < ar.end ? + PAD_AVAIL_OUT + (ar.end - ar.out) : + PAD_AVAIL_OUT - (ar.out - ar.end)); + state->hold = (unsigned long)ar.hold; + state->bits = ar.bits; + return; +} + +#endif diff --git a/contrib/masmx64/inffasx64.asm b/contrib/masmx64/inffasx64.asm new file mode 100644 index 0000000..60a8d89 --- /dev/null +++ b/contrib/masmx64/inffasx64.asm @@ -0,0 +1,396 @@ +; inffasx64.asm is a hand tuned assembler version of inffast.c - fast decoding +; version for AMD64 on Windows using Microsoft C compiler +; +; inffasx64.asm is automatically convert from AMD64 portion of inffas86.c +; inffasx64.asm is called by inffas8664.c, which contain more info. + + +; to compile this file, I use option +; ml64.exe /Flinffasx64 /c /Zi inffasx64.asm +; with Microsoft Macro Assembler (x64) for AMD64 +; + +; This file compile with Microsoft Macro Assembler (x64) for AMD64 +; +; ml64.exe is given with Visual Studio 2005/2008/2010 and Windows WDK +; +; (you can get Windows WDK with ml64 for AMD64 from +; http://www.microsoft.com/whdc/Devtools/wdk/default.mspx for low price) +; + + +.code +inffas8664fnc PROC + +; 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. + + + mov [rsp-8],rsi + mov [rsp-16],rdi + mov [rsp-24],r12 + mov [rsp-32],r13 + mov [rsp-40],r14 + mov [rsp-48],r15 + mov [rsp-56],rbx + + mov rax,rcx + + mov [rax+8], rbp ; /* save regs rbp and rsp */ + mov [rax], rsp + + mov rsp, rax ; /* make rsp point to &ar */ + + mov rsi, [rsp+16] ; /* rsi = in */ + mov rdi, [rsp+32] ; /* rdi = out */ + mov r9, [rsp+24] ; /* r9 = last */ + mov r10, [rsp+48] ; /* r10 = end */ + mov rbp, [rsp+64] ; /* rbp = lcode */ + mov r11, [rsp+72] ; /* r11 = dcode */ + mov rdx, [rsp+80] ; /* rdx = hold */ + mov ebx, [rsp+88] ; /* ebx = bits */ + mov r12d, [rsp+100] ; /* r12d = lmask */ + mov r13d, [rsp+104] ; /* r13d = dmask */ + ; /* r14d = len */ + ; /* r15d = dist */ + + + cld + cmp r10, rdi + je L_one_time ; /* if only one decode left */ + cmp r9, rsi + + jne L_do_loop + + +L_one_time: + mov r8, r12 ; /* r8 = lmask */ + cmp bl, 32 + ja L_get_length_code_one_time + + lodsd ; /* eax = *(uint *)in++ */ + mov cl, bl ; /* cl = bits, needs it for shifting */ + add bl, 32 ; /* bits += 32 */ + shl rax, cl + or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ + jmp L_get_length_code_one_time + +ALIGN 4 +L_while_test: + cmp r10, rdi + jbe L_break_loop + cmp r9, rsi + jbe L_break_loop + +L_do_loop: + mov r8, r12 ; /* r8 = lmask */ + cmp bl, 32 + ja L_get_length_code ; /* if (32 < bits) */ + + lodsd ; /* eax = *(uint *)in++ */ + mov cl, bl ; /* cl = bits, needs it for shifting */ + add bl, 32 ; /* bits += 32 */ + shl rax, cl + or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ + +L_get_length_code: + and r8, rdx ; /* r8 &= hold */ + mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */ + + mov cl, ah ; /* cl = this.bits */ + sub bl, ah ; /* bits -= this.bits */ + shr rdx, cl ; /* hold >>= this.bits */ + + test al, al + jnz L_test_for_length_base ; /* if (op != 0) 45.7% */ + + mov r8, r12 ; /* r8 = lmask */ + shr eax, 16 ; /* output this.val char */ + stosb + +L_get_length_code_one_time: + and r8, rdx ; /* r8 &= hold */ + mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */ + +L_dolen: + mov cl, ah ; /* cl = this.bits */ + sub bl, ah ; /* bits -= this.bits */ + shr rdx, cl ; /* hold >>= this.bits */ + + test al, al + jnz L_test_for_length_base ; /* if (op != 0) 45.7% */ + + shr eax, 16 ; /* output this.val char */ + stosb + jmp L_while_test + +ALIGN 4 +L_test_for_length_base: + mov r14d, eax ; /* len = this */ + shr r14d, 16 ; /* len = this.val */ + mov cl, al + + test al, 16 + jz L_test_for_second_level_length ; /* if ((op & 16) == 0) 8% */ + and cl, 15 ; /* op &= 15 */ + jz L_decode_distance ; /* if (!op) */ + +L_add_bits_to_len: + sub bl, cl + xor eax, eax + inc eax + shl eax, cl + dec eax + and eax, edx ; /* eax &= hold */ + shr rdx, cl + add r14d, eax ; /* len += hold & mask[op] */ + +L_decode_distance: + mov r8, r13 ; /* r8 = dmask */ + cmp bl, 32 + ja L_get_distance_code ; /* if (32 < bits) */ + + lodsd ; /* eax = *(uint *)in++ */ + mov cl, bl ; /* cl = bits, needs it for shifting */ + add bl, 32 ; /* bits += 32 */ + shl rax, cl + or rdx, rax ; /* hold |= *((uint *)in)++ << bits */ + +L_get_distance_code: + and r8, rdx ; /* r8 &= hold */ + mov eax, [r11+r8*4] ; /* eax = dcode[hold & dmask] */ + +L_dodist: + mov r15d, eax ; /* dist = this */ + shr r15d, 16 ; /* dist = this.val */ + mov cl, ah + sub bl, ah ; /* bits -= this.bits */ + shr rdx, cl ; /* hold >>= this.bits */ + mov cl, al ; /* cl = this.op */ + + test al, 16 ; /* if ((op & 16) == 0) */ + jz L_test_for_second_level_dist + and cl, 15 ; /* op &= 15 */ + jz L_check_dist_one + +L_add_bits_to_dist: + sub bl, cl + xor eax, eax + inc eax + shl eax, cl + dec eax ; /* (1 << op) - 1 */ + and eax, edx ; /* eax &= hold */ + shr rdx, cl + add r15d, eax ; /* dist += hold & ((1 << op) - 1) */ + +L_check_window: + mov r8, rsi ; /* save in so from can use it's reg */ + mov rax, rdi + sub rax, [rsp+40] ; /* nbytes = out - beg */ + + cmp eax, r15d + jb L_clip_window ; /* if (dist > nbytes) 4.2% */ + + mov ecx, r14d ; /* ecx = len */ + mov rsi, rdi + sub rsi, r15 ; /* from = out - dist */ + + sar ecx, 1 + jnc L_copy_two ; /* if len % 2 == 0 */ + + rep movsw + mov al, [rsi] + mov [rdi], al + inc rdi + + mov rsi, r8 ; /* move in back to %rsi, toss from */ + jmp L_while_test + +L_copy_two: + rep movsw + mov rsi, r8 ; /* move in back to %rsi, toss from */ + jmp L_while_test + +ALIGN 4 +L_check_dist_one: + cmp r15d, 1 ; /* if dist 1, is a memset */ + jne L_check_window + cmp [rsp+40], rdi ; /* if out == beg, outside window */ + je L_check_window + + mov ecx, r14d ; /* ecx = len */ + mov al, [rdi-1] + mov ah, al + + sar ecx, 1 + jnc L_set_two + mov [rdi], al + inc rdi + +L_set_two: + rep stosw + jmp L_while_test + +ALIGN 4 +L_test_for_second_level_length: + test al, 64 + jnz L_test_for_end_of_block ; /* if ((op & 64) != 0) */ + + xor eax, eax + inc eax + shl eax, cl + dec eax + and eax, edx ; /* eax &= hold */ + add eax, r14d ; /* eax += len */ + mov eax, [rbp+rax*4] ; /* eax = lcode[val+(hold&mask[op])]*/ + jmp L_dolen + +ALIGN 4 +L_test_for_second_level_dist: + test al, 64 + jnz L_invalid_distance_code ; /* if ((op & 64) != 0) */ + + xor eax, eax + inc eax + shl eax, cl + dec eax + and eax, edx ; /* eax &= hold */ + add eax, r15d ; /* eax += dist */ + mov eax, [r11+rax*4] ; /* eax = dcode[val+(hold&mask[op])]*/ + jmp L_dodist + +ALIGN 4 +L_clip_window: + mov ecx, eax ; /* ecx = nbytes */ + mov eax, [rsp+92] ; /* eax = wsize, prepare for dist cmp */ + neg ecx ; /* nbytes = -nbytes */ + + cmp eax, r15d + jb L_invalid_distance_too_far ; /* if (dist > wsize) */ + + add ecx, r15d ; /* nbytes = dist - nbytes */ + cmp dword ptr [rsp+96], 0 + jne L_wrap_around_window ; /* if (write != 0) */ + + mov rsi, [rsp+56] ; /* from = window */ + sub eax, ecx ; /* eax -= nbytes */ + add rsi, rax ; /* from += wsize - nbytes */ + + mov eax, r14d ; /* eax = len */ + cmp r14d, ecx + jbe L_do_copy ; /* if (nbytes >= len) */ + + sub eax, ecx ; /* eax -= nbytes */ + rep movsb + mov rsi, rdi + sub rsi, r15 ; /* from = &out[ -dist ] */ + jmp L_do_copy + +ALIGN 4 +L_wrap_around_window: + mov eax, [rsp+96] ; /* eax = write */ + cmp ecx, eax + jbe L_contiguous_in_window ; /* if (write >= nbytes) */ + + mov esi, [rsp+92] ; /* from = wsize */ + add rsi, [rsp+56] ; /* from += window */ + add rsi, rax ; /* from += write */ + sub rsi, rcx ; /* from -= nbytes */ + sub ecx, eax ; /* nbytes -= write */ + + mov eax, r14d ; /* eax = len */ + cmp eax, ecx + jbe L_do_copy ; /* if (nbytes >= len) */ + + sub eax, ecx ; /* len -= nbytes */ + rep movsb + mov rsi, [rsp+56] ; /* from = window */ + mov ecx, [rsp+96] ; /* nbytes = write */ + cmp eax, ecx + jbe L_do_copy ; /* if (nbytes >= len) */ + + sub eax, ecx ; /* len -= nbytes */ + rep movsb + mov rsi, rdi + sub rsi, r15 ; /* from = out - dist */ + jmp L_do_copy + +ALIGN 4 +L_contiguous_in_window: + mov rsi, [rsp+56] ; /* rsi = window */ + add rsi, rax + sub rsi, rcx ; /* from += write - nbytes */ + + mov eax, r14d ; /* eax = len */ + cmp eax, ecx + jbe L_do_copy ; /* if (nbytes >= len) */ + + sub eax, ecx ; /* len -= nbytes */ + rep movsb + mov rsi, rdi + sub rsi, r15 ; /* from = out - dist */ + jmp L_do_copy ; /* if (nbytes >= len) */ + +ALIGN 4 +L_do_copy: + mov ecx, eax ; /* ecx = len */ + rep movsb + + mov rsi, r8 ; /* move in back to %esi, toss from */ + jmp L_while_test + +L_test_for_end_of_block: + test al, 32 + jz L_invalid_literal_length_code + mov dword ptr [rsp+116], 1 + jmp L_break_loop_with_status + +L_invalid_literal_length_code: + mov dword ptr [rsp+116], 2 + jmp L_break_loop_with_status + +L_invalid_distance_code: + mov dword ptr [rsp+116], 3 + jmp L_break_loop_with_status + +L_invalid_distance_too_far: + mov dword ptr [rsp+116], 4 + jmp L_break_loop_with_status + +L_break_loop: + mov dword ptr [rsp+116], 0 + +L_break_loop_with_status: +; /* put in, out, bits, and hold back into ar and pop esp */ + mov [rsp+16], rsi ; /* in */ + mov [rsp+32], rdi ; /* out */ + mov [rsp+88], ebx ; /* bits */ + mov [rsp+80], rdx ; /* hold */ + + mov rax, [rsp] ; /* restore rbp and rsp */ + mov rbp, [rsp+8] + mov rsp, rax + + + + mov rsi,[rsp-8] + mov rdi,[rsp-16] + mov r12,[rsp-24] + mov r13,[rsp-32] + mov r14,[rsp-40] + mov r15,[rsp-48] + mov rbx,[rsp-56] + + ret 0 +; : +; : "m" (ar) +; : "memory", "%rax", "%rbx", "%rcx", "%rdx", "%rsi", "%rdi", +; "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15" +; ); + +inffas8664fnc ENDP +;_TEXT ENDS +END diff --git a/contrib/masmx64/readme.txt b/contrib/masmx64/readme.txt new file mode 100644 index 0000000..2da6733 --- /dev/null +++ b/contrib/masmx64/readme.txt @@ -0,0 +1,31 @@ +Summary +------- +This directory contains ASM implementations of the functions +longest_match() and inflate_fast(), for 64 bits x86 (both AMD64 and Intel EM64t), +for use with Microsoft Macro Assembler (x64) for AMD64 and Microsoft C++ 64 bits. + +gvmat64.asm is written by Gilles Vollant (2005), by using Brian Raiter 686/32 bits + assembly optimized version from Jean-loup Gailly original longest_match function + +inffasx64.asm and inffas8664.c were written by Chris Anderson, by optimizing + original function from Mark Adler + +Use instructions +---------------- +Assemble the .asm files using MASM and put the object files into the zlib source +directory. You can also get object files here: + + http://www.winimage.com/zLibDll/zlib124_masm_obj.zip + +define ASMV and ASMINF in your project. Include inffas8664.c in your source tree, +and inffasx64.obj and gvmat64.obj as object to link. + + +Build instructions +------------------ +run bld_64.bat with Microsoft Macro Assembler (x64) for AMD64 (ml64.exe) + +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) -- cgit v1.1