1/* match.S -- x86 assembly version of the zlib longest_match() function. 2 * Optimized for the Intel 686 chips (PPro and later). 3 * 4 * Copyright (C) 1998, 2007 Brian Raiter <breadbox@muppetlabs.com> 5 * 6 * This software is provided 'as-is', without any express or implied 7 * warranty. In no event will the author be held liable for any damages 8 * arising from the use of this software. 9 * 10 * Permission is granted to anyone to use this software for any purpose, 11 * including commercial applications, and to alter it and redistribute it 12 * freely, subject to the following restrictions: 13 * 14 * 1. The origin of this software must not be misrepresented; you must not 15 * claim that you wrote the original software. If you use this software 16 * in a product, an acknowledgment in the product documentation would be 17 * appreciated but is not required. 18 * 2. Altered source versions must be plainly marked as such, and must not be 19 * misrepresented as being the original software. 20 * 3. This notice may not be removed or altered from any source distribution. 21 */ 22 23#ifndef NO_UNDERLINE 24#define match_init _match_init 25#define longest_match _longest_match 26#endif 27 28#define MAX_MATCH (258) 29#define MIN_MATCH (3) 30#define MIN_LOOKAHEAD (MAX_MATCH + MIN_MATCH + 1) 31#define MAX_MATCH_8 ((MAX_MATCH + 7) & ~7) 32 33/* stack frame offsets */ 34 35#define chainlenwmask 0 /* high word: current chain len */ 36 /* low word: s->wmask */ 37#define window 4 /* local copy of s->window */ 38#define windowbestlen 8 /* s->window + bestlen */ 39#define scanstart 16 /* first two bytes of string */ 40#define scanend 12 /* last two bytes of string */ 41#define scanalign 20 /* dword-misalignment of string */ 42#define nicematch 24 /* a good enough match size */ 43#define bestlen 28 /* size of best match so far */ 44#define scan 32 /* ptr to string wanting match */ 45 46#define LocalVarsSize (36) 47/* saved ebx 36 */ 48/* saved edi 40 */ 49/* saved esi 44 */ 50/* saved ebp 48 */ 51/* return address 52 */ 52#define deflatestate 56 /* the function arguments */ 53#define curmatch 60 54 55/* All the +zlib1222add offsets are due to the addition of fields 56 * in zlib in the deflate_state structure since the asm code was first written 57 * (if you compile with zlib 1.0.4 or older, use "zlib1222add equ (-4)"). 58 * (if you compile with zlib between 1.0.5 and 1.2.2.1, use "zlib1222add equ 0"). 59 * if you compile with zlib 1.2.2.2 or later , use "zlib1222add equ 8"). 60 */ 61 62#define zlib1222add (8) 63 64#define dsWSize (36+zlib1222add) 65#define dsWMask (44+zlib1222add) 66#define dsWindow (48+zlib1222add) 67#define dsPrev (56+zlib1222add) 68#define dsMatchLen (88+zlib1222add) 69#define dsPrevMatch (92+zlib1222add) 70#define dsStrStart (100+zlib1222add) 71#define dsMatchStart (104+zlib1222add) 72#define dsLookahead (108+zlib1222add) 73#define dsPrevLen (112+zlib1222add) 74#define dsMaxChainLen (116+zlib1222add) 75#define dsGoodMatch (132+zlib1222add) 76#define dsNiceMatch (136+zlib1222add) 77 78 79.file "match.S" 80 81.globl match_init, longest_match 82 83.text 84 85/* uInt longest_match(deflate_state *deflatestate, IPos curmatch) */ 86 87longest_match: 88 89/* Save registers that the compiler may be using, and adjust %esp to */ 90/* make room for our stack frame. */ 91 92 pushl %ebp 93 pushl %edi 94 pushl %esi 95 pushl %ebx 96 subl $LocalVarsSize, %esp 97 98/* Retrieve the function arguments. %ecx will hold cur_match */ 99/* throughout the entire function. %edx will hold the pointer to the */ 100/* deflate_state structure during the function's setup (before */ 101/* entering the main loop). */ 102 103 movl deflatestate(%esp), %edx 104 movl curmatch(%esp), %ecx 105 106/* uInt wmask = s->w_mask; */ 107/* unsigned chain_length = s->max_chain_length; */ 108/* if (s->prev_length >= s->good_match) { */ 109/* chain_length >>= 2; */ 110/* } */ 111 112 movl dsPrevLen(%edx), %eax 113 movl dsGoodMatch(%edx), %ebx 114 cmpl %ebx, %eax 115 movl dsWMask(%edx), %eax 116 movl dsMaxChainLen(%edx), %ebx 117 jl LastMatchGood 118 shrl $2, %ebx 119LastMatchGood: 120 121/* chainlen is decremented once beforehand so that the function can */ 122/* use the sign flag instead of the zero flag for the exit test. */ 123/* It is then shifted into the high word, to make room for the wmask */ 124/* value, which it will always accompany. */ 125 126 decl %ebx 127 shll $16, %ebx 128 orl %eax, %ebx 129 movl %ebx, chainlenwmask(%esp) 130 131/* if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead; */ 132 133 movl dsNiceMatch(%edx), %eax 134 movl dsLookahead(%edx), %ebx 135 cmpl %eax, %ebx 136 jl LookaheadLess 137 movl %eax, %ebx 138LookaheadLess: movl %ebx, nicematch(%esp) 139 140/* register Bytef *scan = s->window + s->strstart; */ 141 142 movl dsWindow(%edx), %esi 143 movl %esi, window(%esp) 144 movl dsStrStart(%edx), %ebp 145 lea (%esi,%ebp), %edi 146 movl %edi, scan(%esp) 147 148/* Determine how many bytes the scan ptr is off from being */ 149/* dword-aligned. */ 150 151 movl %edi, %eax 152 negl %eax 153 andl $3, %eax 154 movl %eax, scanalign(%esp) 155 156/* IPos limit = s->strstart > (IPos)MAX_DIST(s) ? */ 157/* s->strstart - (IPos)MAX_DIST(s) : NIL; */ 158 159 movl dsWSize(%edx), %eax 160 subl $MIN_LOOKAHEAD, %eax 161 subl %eax, %ebp 162 jg LimitPositive 163 xorl %ebp, %ebp 164LimitPositive: 165 166/* int best_len = s->prev_length; */ 167 168 movl dsPrevLen(%edx), %eax 169 movl %eax, bestlen(%esp) 170 171/* Store the sum of s->window + best_len in %esi locally, and in %esi. */ 172 173 addl %eax, %esi 174 movl %esi, windowbestlen(%esp) 175 176/* register ush scan_start = *(ushf*)scan; */ 177/* register ush scan_end = *(ushf*)(scan+best_len-1); */ 178/* Posf *prev = s->prev; */ 179 180 movzwl (%edi), %ebx 181 movl %ebx, scanstart(%esp) 182 movzwl -1(%edi,%eax), %ebx 183 movl %ebx, scanend(%esp) 184 movl dsPrev(%edx), %edi 185 186/* Jump into the main loop. */ 187 188 movl chainlenwmask(%esp), %edx 189 jmp LoopEntry 190 191.balign 16 192 193/* do { 194 * match = s->window + cur_match; 195 * if (*(ushf*)(match+best_len-1) != scan_end || 196 * *(ushf*)match != scan_start) continue; 197 * [...] 198 * } while ((cur_match = prev[cur_match & wmask]) > limit 199 * && --chain_length != 0); 200 * 201 * Here is the inner loop of the function. The function will spend the 202 * majority of its time in this loop, and majority of that time will 203 * be spent in the first ten instructions. 204 * 205 * Within this loop: 206 * %ebx = scanend 207 * %ecx = curmatch 208 * %edx = chainlenwmask - i.e., ((chainlen << 16) | wmask) 209 * %esi = windowbestlen - i.e., (window + bestlen) 210 * %edi = prev 211 * %ebp = limit 212 */ 213LookupLoop: 214 andl %edx, %ecx 215 movzwl (%edi,%ecx,2), %ecx 216 cmpl %ebp, %ecx 217 jbe LeaveNow 218 subl $0x00010000, %edx 219 js LeaveNow 220LoopEntry: movzwl -1(%esi,%ecx), %eax 221 cmpl %ebx, %eax 222 jnz LookupLoop 223 movl window(%esp), %eax 224 movzwl (%eax,%ecx), %eax 225 cmpl scanstart(%esp), %eax 226 jnz LookupLoop 227 228/* Store the current value of chainlen. */ 229 230 movl %edx, chainlenwmask(%esp) 231 232/* Point %edi to the string under scrutiny, and %esi to the string we */ 233/* are hoping to match it up with. In actuality, %esi and %edi are */ 234/* both pointed (MAX_MATCH_8 - scanalign) bytes ahead, and %edx is */ 235/* initialized to -(MAX_MATCH_8 - scanalign). */ 236 237 movl window(%esp), %esi 238 movl scan(%esp), %edi 239 addl %ecx, %esi 240 movl scanalign(%esp), %eax 241 movl $(-MAX_MATCH_8), %edx 242 lea MAX_MATCH_8(%edi,%eax), %edi 243 lea MAX_MATCH_8(%esi,%eax), %esi 244 245/* Test the strings for equality, 8 bytes at a time. At the end, 246 * adjust %edx so that it is offset to the exact byte that mismatched. 247 * 248 * We already know at this point that the first three bytes of the 249 * strings match each other, and they can be safely passed over before 250 * starting the compare loop. So what this code does is skip over 0-3 251 * bytes, as much as necessary in order to dword-align the %edi 252 * pointer. (%esi will still be misaligned three times out of four.) 253 * 254 * It should be confessed that this loop usually does not represent 255 * much of the total running time. Replacing it with a more 256 * straightforward "rep cmpsb" would not drastically degrade 257 * performance. 258 */ 259LoopCmps: 260 movl (%esi,%edx), %eax 261 xorl (%edi,%edx), %eax 262 jnz LeaveLoopCmps 263 movl 4(%esi,%edx), %eax 264 xorl 4(%edi,%edx), %eax 265 jnz LeaveLoopCmps4 266 addl $8, %edx 267 jnz LoopCmps 268 jmp LenMaximum 269LeaveLoopCmps4: addl $4, %edx 270LeaveLoopCmps: testl $0x0000FFFF, %eax 271 jnz LenLower 272 addl $2, %edx 273 shrl $16, %eax 274LenLower: subb $1, %al 275 adcl $0, %edx 276 277/* Calculate the length of the match. If it is longer than MAX_MATCH, */ 278/* then automatically accept it as the best possible match and leave. */ 279 280 lea (%edi,%edx), %eax 281 movl scan(%esp), %edi 282 subl %edi, %eax 283 cmpl $MAX_MATCH, %eax 284 jge LenMaximum 285 286/* If the length of the match is not longer than the best match we */ 287/* have so far, then forget it and return to the lookup loop. */ 288 289 movl deflatestate(%esp), %edx 290 movl bestlen(%esp), %ebx 291 cmpl %ebx, %eax 292 jg LongerMatch 293 movl windowbestlen(%esp), %esi 294 movl dsPrev(%edx), %edi 295 movl scanend(%esp), %ebx 296 movl chainlenwmask(%esp), %edx 297 jmp LookupLoop 298 299/* s->match_start = cur_match; */ 300/* best_len = len; */ 301/* if (len >= nice_match) break; */ 302/* scan_end = *(ushf*)(scan+best_len-1); */ 303 304LongerMatch: movl nicematch(%esp), %ebx 305 movl %eax, bestlen(%esp) 306 movl %ecx, dsMatchStart(%edx) 307 cmpl %ebx, %eax 308 jge LeaveNow 309 movl window(%esp), %esi 310 addl %eax, %esi 311 movl %esi, windowbestlen(%esp) 312 movzwl -1(%edi,%eax), %ebx 313 movl dsPrev(%edx), %edi 314 movl %ebx, scanend(%esp) 315 movl chainlenwmask(%esp), %edx 316 jmp LookupLoop 317 318/* Accept the current string, with the maximum possible length. */ 319 320LenMaximum: movl deflatestate(%esp), %edx 321 movl $MAX_MATCH, bestlen(%esp) 322 movl %ecx, dsMatchStart(%edx) 323 324/* if ((uInt)best_len <= s->lookahead) return (uInt)best_len; */ 325/* return s->lookahead; */ 326 327LeaveNow: 328 movl deflatestate(%esp), %edx 329 movl bestlen(%esp), %ebx 330 movl dsLookahead(%edx), %eax 331 cmpl %eax, %ebx 332 jg LookaheadRet 333 movl %ebx, %eax 334LookaheadRet: 335 336/* Restore the stack and return from whence we came. */ 337 338 addl $LocalVarsSize, %esp 339 popl %ebx 340 popl %esi 341 popl %edi 342 popl %ebp 343match_init: ret 344