Chris@4
|
1 ; inffasx64.asm is a hand tuned assembler version of inffast.c - fast decoding
|
Chris@4
|
2 ; version for AMD64 on Windows using Microsoft C compiler
|
Chris@4
|
3 ;
|
Chris@4
|
4 ; inffasx64.asm is automatically convert from AMD64 portion of inffas86.c
|
Chris@4
|
5 ; inffasx64.asm is called by inffas8664.c, which contain more info.
|
Chris@4
|
6
|
Chris@4
|
7
|
Chris@4
|
8 ; to compile this file, I use option
|
Chris@4
|
9 ; ml64.exe /Flinffasx64 /c /Zi inffasx64.asm
|
Chris@4
|
10 ; with Microsoft Macro Assembler (x64) for AMD64
|
Chris@4
|
11 ;
|
Chris@4
|
12
|
Chris@4
|
13 ; This file compile with Microsoft Macro Assembler (x64) for AMD64
|
Chris@4
|
14 ;
|
Chris@4
|
15 ; ml64.exe is given with Visual Studio 2005/2008/2010 and Windows WDK
|
Chris@4
|
16 ;
|
Chris@4
|
17 ; (you can get Windows WDK with ml64 for AMD64 from
|
Chris@4
|
18 ; http://www.microsoft.com/whdc/Devtools/wdk/default.mspx for low price)
|
Chris@4
|
19 ;
|
Chris@4
|
20
|
Chris@4
|
21
|
Chris@4
|
22 .code
|
Chris@4
|
23 inffas8664fnc PROC
|
Chris@4
|
24
|
Chris@4
|
25 ; see http://weblogs.asp.net/oldnewthing/archive/2004/01/14/58579.aspx and
|
Chris@4
|
26 ; http://msdn.microsoft.com/library/en-us/kmarch/hh/kmarch/64bitAMD_8e951dd2-ee77-4728-8702-55ce4b5dd24a.xml.asp
|
Chris@4
|
27 ;
|
Chris@4
|
28 ; All registers must be preserved across the call, except for
|
Chris@4
|
29 ; rax, rcx, rdx, r8, r-9, r10, and r11, which are scratch.
|
Chris@4
|
30
|
Chris@4
|
31
|
Chris@4
|
32 mov [rsp-8],rsi
|
Chris@4
|
33 mov [rsp-16],rdi
|
Chris@4
|
34 mov [rsp-24],r12
|
Chris@4
|
35 mov [rsp-32],r13
|
Chris@4
|
36 mov [rsp-40],r14
|
Chris@4
|
37 mov [rsp-48],r15
|
Chris@4
|
38 mov [rsp-56],rbx
|
Chris@4
|
39
|
Chris@4
|
40 mov rax,rcx
|
Chris@4
|
41
|
Chris@4
|
42 mov [rax+8], rbp ; /* save regs rbp and rsp */
|
Chris@4
|
43 mov [rax], rsp
|
Chris@4
|
44
|
Chris@4
|
45 mov rsp, rax ; /* make rsp point to &ar */
|
Chris@4
|
46
|
Chris@4
|
47 mov rsi, [rsp+16] ; /* rsi = in */
|
Chris@4
|
48 mov rdi, [rsp+32] ; /* rdi = out */
|
Chris@4
|
49 mov r9, [rsp+24] ; /* r9 = last */
|
Chris@4
|
50 mov r10, [rsp+48] ; /* r10 = end */
|
Chris@4
|
51 mov rbp, [rsp+64] ; /* rbp = lcode */
|
Chris@4
|
52 mov r11, [rsp+72] ; /* r11 = dcode */
|
Chris@4
|
53 mov rdx, [rsp+80] ; /* rdx = hold */
|
Chris@4
|
54 mov ebx, [rsp+88] ; /* ebx = bits */
|
Chris@4
|
55 mov r12d, [rsp+100] ; /* r12d = lmask */
|
Chris@4
|
56 mov r13d, [rsp+104] ; /* r13d = dmask */
|
Chris@4
|
57 ; /* r14d = len */
|
Chris@4
|
58 ; /* r15d = dist */
|
Chris@4
|
59
|
Chris@4
|
60
|
Chris@4
|
61 cld
|
Chris@4
|
62 cmp r10, rdi
|
Chris@4
|
63 je L_one_time ; /* if only one decode left */
|
Chris@4
|
64 cmp r9, rsi
|
Chris@4
|
65
|
Chris@4
|
66 jne L_do_loop
|
Chris@4
|
67
|
Chris@4
|
68
|
Chris@4
|
69 L_one_time:
|
Chris@4
|
70 mov r8, r12 ; /* r8 = lmask */
|
Chris@4
|
71 cmp bl, 32
|
Chris@4
|
72 ja L_get_length_code_one_time
|
Chris@4
|
73
|
Chris@4
|
74 lodsd ; /* eax = *(uint *)in++ */
|
Chris@4
|
75 mov cl, bl ; /* cl = bits, needs it for shifting */
|
Chris@4
|
76 add bl, 32 ; /* bits += 32 */
|
Chris@4
|
77 shl rax, cl
|
Chris@4
|
78 or rdx, rax ; /* hold |= *((uint *)in)++ << bits */
|
Chris@4
|
79 jmp L_get_length_code_one_time
|
Chris@4
|
80
|
Chris@4
|
81 ALIGN 4
|
Chris@4
|
82 L_while_test:
|
Chris@4
|
83 cmp r10, rdi
|
Chris@4
|
84 jbe L_break_loop
|
Chris@4
|
85 cmp r9, rsi
|
Chris@4
|
86 jbe L_break_loop
|
Chris@4
|
87
|
Chris@4
|
88 L_do_loop:
|
Chris@4
|
89 mov r8, r12 ; /* r8 = lmask */
|
Chris@4
|
90 cmp bl, 32
|
Chris@4
|
91 ja L_get_length_code ; /* if (32 < bits) */
|
Chris@4
|
92
|
Chris@4
|
93 lodsd ; /* eax = *(uint *)in++ */
|
Chris@4
|
94 mov cl, bl ; /* cl = bits, needs it for shifting */
|
Chris@4
|
95 add bl, 32 ; /* bits += 32 */
|
Chris@4
|
96 shl rax, cl
|
Chris@4
|
97 or rdx, rax ; /* hold |= *((uint *)in)++ << bits */
|
Chris@4
|
98
|
Chris@4
|
99 L_get_length_code:
|
Chris@4
|
100 and r8, rdx ; /* r8 &= hold */
|
Chris@4
|
101 mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */
|
Chris@4
|
102
|
Chris@4
|
103 mov cl, ah ; /* cl = this.bits */
|
Chris@4
|
104 sub bl, ah ; /* bits -= this.bits */
|
Chris@4
|
105 shr rdx, cl ; /* hold >>= this.bits */
|
Chris@4
|
106
|
Chris@4
|
107 test al, al
|
Chris@4
|
108 jnz L_test_for_length_base ; /* if (op != 0) 45.7% */
|
Chris@4
|
109
|
Chris@4
|
110 mov r8, r12 ; /* r8 = lmask */
|
Chris@4
|
111 shr eax, 16 ; /* output this.val char */
|
Chris@4
|
112 stosb
|
Chris@4
|
113
|
Chris@4
|
114 L_get_length_code_one_time:
|
Chris@4
|
115 and r8, rdx ; /* r8 &= hold */
|
Chris@4
|
116 mov eax, [rbp+r8*4] ; /* eax = lcode[hold & lmask] */
|
Chris@4
|
117
|
Chris@4
|
118 L_dolen:
|
Chris@4
|
119 mov cl, ah ; /* cl = this.bits */
|
Chris@4
|
120 sub bl, ah ; /* bits -= this.bits */
|
Chris@4
|
121 shr rdx, cl ; /* hold >>= this.bits */
|
Chris@4
|
122
|
Chris@4
|
123 test al, al
|
Chris@4
|
124 jnz L_test_for_length_base ; /* if (op != 0) 45.7% */
|
Chris@4
|
125
|
Chris@4
|
126 shr eax, 16 ; /* output this.val char */
|
Chris@4
|
127 stosb
|
Chris@4
|
128 jmp L_while_test
|
Chris@4
|
129
|
Chris@4
|
130 ALIGN 4
|
Chris@4
|
131 L_test_for_length_base:
|
Chris@4
|
132 mov r14d, eax ; /* len = this */
|
Chris@4
|
133 shr r14d, 16 ; /* len = this.val */
|
Chris@4
|
134 mov cl, al
|
Chris@4
|
135
|
Chris@4
|
136 test al, 16
|
Chris@4
|
137 jz L_test_for_second_level_length ; /* if ((op & 16) == 0) 8% */
|
Chris@4
|
138 and cl, 15 ; /* op &= 15 */
|
Chris@4
|
139 jz L_decode_distance ; /* if (!op) */
|
Chris@4
|
140
|
Chris@4
|
141 L_add_bits_to_len:
|
Chris@4
|
142 sub bl, cl
|
Chris@4
|
143 xor eax, eax
|
Chris@4
|
144 inc eax
|
Chris@4
|
145 shl eax, cl
|
Chris@4
|
146 dec eax
|
Chris@4
|
147 and eax, edx ; /* eax &= hold */
|
Chris@4
|
148 shr rdx, cl
|
Chris@4
|
149 add r14d, eax ; /* len += hold & mask[op] */
|
Chris@4
|
150
|
Chris@4
|
151 L_decode_distance:
|
Chris@4
|
152 mov r8, r13 ; /* r8 = dmask */
|
Chris@4
|
153 cmp bl, 32
|
Chris@4
|
154 ja L_get_distance_code ; /* if (32 < bits) */
|
Chris@4
|
155
|
Chris@4
|
156 lodsd ; /* eax = *(uint *)in++ */
|
Chris@4
|
157 mov cl, bl ; /* cl = bits, needs it for shifting */
|
Chris@4
|
158 add bl, 32 ; /* bits += 32 */
|
Chris@4
|
159 shl rax, cl
|
Chris@4
|
160 or rdx, rax ; /* hold |= *((uint *)in)++ << bits */
|
Chris@4
|
161
|
Chris@4
|
162 L_get_distance_code:
|
Chris@4
|
163 and r8, rdx ; /* r8 &= hold */
|
Chris@4
|
164 mov eax, [r11+r8*4] ; /* eax = dcode[hold & dmask] */
|
Chris@4
|
165
|
Chris@4
|
166 L_dodist:
|
Chris@4
|
167 mov r15d, eax ; /* dist = this */
|
Chris@4
|
168 shr r15d, 16 ; /* dist = this.val */
|
Chris@4
|
169 mov cl, ah
|
Chris@4
|
170 sub bl, ah ; /* bits -= this.bits */
|
Chris@4
|
171 shr rdx, cl ; /* hold >>= this.bits */
|
Chris@4
|
172 mov cl, al ; /* cl = this.op */
|
Chris@4
|
173
|
Chris@4
|
174 test al, 16 ; /* if ((op & 16) == 0) */
|
Chris@4
|
175 jz L_test_for_second_level_dist
|
Chris@4
|
176 and cl, 15 ; /* op &= 15 */
|
Chris@4
|
177 jz L_check_dist_one
|
Chris@4
|
178
|
Chris@4
|
179 L_add_bits_to_dist:
|
Chris@4
|
180 sub bl, cl
|
Chris@4
|
181 xor eax, eax
|
Chris@4
|
182 inc eax
|
Chris@4
|
183 shl eax, cl
|
Chris@4
|
184 dec eax ; /* (1 << op) - 1 */
|
Chris@4
|
185 and eax, edx ; /* eax &= hold */
|
Chris@4
|
186 shr rdx, cl
|
Chris@4
|
187 add r15d, eax ; /* dist += hold & ((1 << op) - 1) */
|
Chris@4
|
188
|
Chris@4
|
189 L_check_window:
|
Chris@4
|
190 mov r8, rsi ; /* save in so from can use it's reg */
|
Chris@4
|
191 mov rax, rdi
|
Chris@4
|
192 sub rax, [rsp+40] ; /* nbytes = out - beg */
|
Chris@4
|
193
|
Chris@4
|
194 cmp eax, r15d
|
Chris@4
|
195 jb L_clip_window ; /* if (dist > nbytes) 4.2% */
|
Chris@4
|
196
|
Chris@4
|
197 mov ecx, r14d ; /* ecx = len */
|
Chris@4
|
198 mov rsi, rdi
|
Chris@4
|
199 sub rsi, r15 ; /* from = out - dist */
|
Chris@4
|
200
|
Chris@4
|
201 sar ecx, 1
|
Chris@4
|
202 jnc L_copy_two ; /* if len % 2 == 0 */
|
Chris@4
|
203
|
Chris@4
|
204 rep movsw
|
Chris@4
|
205 mov al, [rsi]
|
Chris@4
|
206 mov [rdi], al
|
Chris@4
|
207 inc rdi
|
Chris@4
|
208
|
Chris@4
|
209 mov rsi, r8 ; /* move in back to %rsi, toss from */
|
Chris@4
|
210 jmp L_while_test
|
Chris@4
|
211
|
Chris@4
|
212 L_copy_two:
|
Chris@4
|
213 rep movsw
|
Chris@4
|
214 mov rsi, r8 ; /* move in back to %rsi, toss from */
|
Chris@4
|
215 jmp L_while_test
|
Chris@4
|
216
|
Chris@4
|
217 ALIGN 4
|
Chris@4
|
218 L_check_dist_one:
|
Chris@4
|
219 cmp r15d, 1 ; /* if dist 1, is a memset */
|
Chris@4
|
220 jne L_check_window
|
Chris@4
|
221 cmp [rsp+40], rdi ; /* if out == beg, outside window */
|
Chris@4
|
222 je L_check_window
|
Chris@4
|
223
|
Chris@4
|
224 mov ecx, r14d ; /* ecx = len */
|
Chris@4
|
225 mov al, [rdi-1]
|
Chris@4
|
226 mov ah, al
|
Chris@4
|
227
|
Chris@4
|
228 sar ecx, 1
|
Chris@4
|
229 jnc L_set_two
|
Chris@4
|
230 mov [rdi], al
|
Chris@4
|
231 inc rdi
|
Chris@4
|
232
|
Chris@4
|
233 L_set_two:
|
Chris@4
|
234 rep stosw
|
Chris@4
|
235 jmp L_while_test
|
Chris@4
|
236
|
Chris@4
|
237 ALIGN 4
|
Chris@4
|
238 L_test_for_second_level_length:
|
Chris@4
|
239 test al, 64
|
Chris@4
|
240 jnz L_test_for_end_of_block ; /* if ((op & 64) != 0) */
|
Chris@4
|
241
|
Chris@4
|
242 xor eax, eax
|
Chris@4
|
243 inc eax
|
Chris@4
|
244 shl eax, cl
|
Chris@4
|
245 dec eax
|
Chris@4
|
246 and eax, edx ; /* eax &= hold */
|
Chris@4
|
247 add eax, r14d ; /* eax += len */
|
Chris@4
|
248 mov eax, [rbp+rax*4] ; /* eax = lcode[val+(hold&mask[op])]*/
|
Chris@4
|
249 jmp L_dolen
|
Chris@4
|
250
|
Chris@4
|
251 ALIGN 4
|
Chris@4
|
252 L_test_for_second_level_dist:
|
Chris@4
|
253 test al, 64
|
Chris@4
|
254 jnz L_invalid_distance_code ; /* if ((op & 64) != 0) */
|
Chris@4
|
255
|
Chris@4
|
256 xor eax, eax
|
Chris@4
|
257 inc eax
|
Chris@4
|
258 shl eax, cl
|
Chris@4
|
259 dec eax
|
Chris@4
|
260 and eax, edx ; /* eax &= hold */
|
Chris@4
|
261 add eax, r15d ; /* eax += dist */
|
Chris@4
|
262 mov eax, [r11+rax*4] ; /* eax = dcode[val+(hold&mask[op])]*/
|
Chris@4
|
263 jmp L_dodist
|
Chris@4
|
264
|
Chris@4
|
265 ALIGN 4
|
Chris@4
|
266 L_clip_window:
|
Chris@4
|
267 mov ecx, eax ; /* ecx = nbytes */
|
Chris@4
|
268 mov eax, [rsp+92] ; /* eax = wsize, prepare for dist cmp */
|
Chris@4
|
269 neg ecx ; /* nbytes = -nbytes */
|
Chris@4
|
270
|
Chris@4
|
271 cmp eax, r15d
|
Chris@4
|
272 jb L_invalid_distance_too_far ; /* if (dist > wsize) */
|
Chris@4
|
273
|
Chris@4
|
274 add ecx, r15d ; /* nbytes = dist - nbytes */
|
Chris@4
|
275 cmp dword ptr [rsp+96], 0
|
Chris@4
|
276 jne L_wrap_around_window ; /* if (write != 0) */
|
Chris@4
|
277
|
Chris@4
|
278 mov rsi, [rsp+56] ; /* from = window */
|
Chris@4
|
279 sub eax, ecx ; /* eax -= nbytes */
|
Chris@4
|
280 add rsi, rax ; /* from += wsize - nbytes */
|
Chris@4
|
281
|
Chris@4
|
282 mov eax, r14d ; /* eax = len */
|
Chris@4
|
283 cmp r14d, ecx
|
Chris@4
|
284 jbe L_do_copy ; /* if (nbytes >= len) */
|
Chris@4
|
285
|
Chris@4
|
286 sub eax, ecx ; /* eax -= nbytes */
|
Chris@4
|
287 rep movsb
|
Chris@4
|
288 mov rsi, rdi
|
Chris@4
|
289 sub rsi, r15 ; /* from = &out[ -dist ] */
|
Chris@4
|
290 jmp L_do_copy
|
Chris@4
|
291
|
Chris@4
|
292 ALIGN 4
|
Chris@4
|
293 L_wrap_around_window:
|
Chris@4
|
294 mov eax, [rsp+96] ; /* eax = write */
|
Chris@4
|
295 cmp ecx, eax
|
Chris@4
|
296 jbe L_contiguous_in_window ; /* if (write >= nbytes) */
|
Chris@4
|
297
|
Chris@4
|
298 mov esi, [rsp+92] ; /* from = wsize */
|
Chris@4
|
299 add rsi, [rsp+56] ; /* from += window */
|
Chris@4
|
300 add rsi, rax ; /* from += write */
|
Chris@4
|
301 sub rsi, rcx ; /* from -= nbytes */
|
Chris@4
|
302 sub ecx, eax ; /* nbytes -= write */
|
Chris@4
|
303
|
Chris@4
|
304 mov eax, r14d ; /* eax = len */
|
Chris@4
|
305 cmp eax, ecx
|
Chris@4
|
306 jbe L_do_copy ; /* if (nbytes >= len) */
|
Chris@4
|
307
|
Chris@4
|
308 sub eax, ecx ; /* len -= nbytes */
|
Chris@4
|
309 rep movsb
|
Chris@4
|
310 mov rsi, [rsp+56] ; /* from = window */
|
Chris@4
|
311 mov ecx, [rsp+96] ; /* nbytes = write */
|
Chris@4
|
312 cmp eax, ecx
|
Chris@4
|
313 jbe L_do_copy ; /* if (nbytes >= len) */
|
Chris@4
|
314
|
Chris@4
|
315 sub eax, ecx ; /* len -= nbytes */
|
Chris@4
|
316 rep movsb
|
Chris@4
|
317 mov rsi, rdi
|
Chris@4
|
318 sub rsi, r15 ; /* from = out - dist */
|
Chris@4
|
319 jmp L_do_copy
|
Chris@4
|
320
|
Chris@4
|
321 ALIGN 4
|
Chris@4
|
322 L_contiguous_in_window:
|
Chris@4
|
323 mov rsi, [rsp+56] ; /* rsi = window */
|
Chris@4
|
324 add rsi, rax
|
Chris@4
|
325 sub rsi, rcx ; /* from += write - nbytes */
|
Chris@4
|
326
|
Chris@4
|
327 mov eax, r14d ; /* eax = len */
|
Chris@4
|
328 cmp eax, ecx
|
Chris@4
|
329 jbe L_do_copy ; /* if (nbytes >= len) */
|
Chris@4
|
330
|
Chris@4
|
331 sub eax, ecx ; /* len -= nbytes */
|
Chris@4
|
332 rep movsb
|
Chris@4
|
333 mov rsi, rdi
|
Chris@4
|
334 sub rsi, r15 ; /* from = out - dist */
|
Chris@4
|
335 jmp L_do_copy ; /* if (nbytes >= len) */
|
Chris@4
|
336
|
Chris@4
|
337 ALIGN 4
|
Chris@4
|
338 L_do_copy:
|
Chris@4
|
339 mov ecx, eax ; /* ecx = len */
|
Chris@4
|
340 rep movsb
|
Chris@4
|
341
|
Chris@4
|
342 mov rsi, r8 ; /* move in back to %esi, toss from */
|
Chris@4
|
343 jmp L_while_test
|
Chris@4
|
344
|
Chris@4
|
345 L_test_for_end_of_block:
|
Chris@4
|
346 test al, 32
|
Chris@4
|
347 jz L_invalid_literal_length_code
|
Chris@4
|
348 mov dword ptr [rsp+116], 1
|
Chris@4
|
349 jmp L_break_loop_with_status
|
Chris@4
|
350
|
Chris@4
|
351 L_invalid_literal_length_code:
|
Chris@4
|
352 mov dword ptr [rsp+116], 2
|
Chris@4
|
353 jmp L_break_loop_with_status
|
Chris@4
|
354
|
Chris@4
|
355 L_invalid_distance_code:
|
Chris@4
|
356 mov dword ptr [rsp+116], 3
|
Chris@4
|
357 jmp L_break_loop_with_status
|
Chris@4
|
358
|
Chris@4
|
359 L_invalid_distance_too_far:
|
Chris@4
|
360 mov dword ptr [rsp+116], 4
|
Chris@4
|
361 jmp L_break_loop_with_status
|
Chris@4
|
362
|
Chris@4
|
363 L_break_loop:
|
Chris@4
|
364 mov dword ptr [rsp+116], 0
|
Chris@4
|
365
|
Chris@4
|
366 L_break_loop_with_status:
|
Chris@4
|
367 ; /* put in, out, bits, and hold back into ar and pop esp */
|
Chris@4
|
368 mov [rsp+16], rsi ; /* in */
|
Chris@4
|
369 mov [rsp+32], rdi ; /* out */
|
Chris@4
|
370 mov [rsp+88], ebx ; /* bits */
|
Chris@4
|
371 mov [rsp+80], rdx ; /* hold */
|
Chris@4
|
372
|
Chris@4
|
373 mov rax, [rsp] ; /* restore rbp and rsp */
|
Chris@4
|
374 mov rbp, [rsp+8]
|
Chris@4
|
375 mov rsp, rax
|
Chris@4
|
376
|
Chris@4
|
377
|
Chris@4
|
378
|
Chris@4
|
379 mov rsi,[rsp-8]
|
Chris@4
|
380 mov rdi,[rsp-16]
|
Chris@4
|
381 mov r12,[rsp-24]
|
Chris@4
|
382 mov r13,[rsp-32]
|
Chris@4
|
383 mov r14,[rsp-40]
|
Chris@4
|
384 mov r15,[rsp-48]
|
Chris@4
|
385 mov rbx,[rsp-56]
|
Chris@4
|
386
|
Chris@4
|
387 ret 0
|
Chris@4
|
388 ; :
|
Chris@4
|
389 ; : "m" (ar)
|
Chris@4
|
390 ; : "memory", "%rax", "%rbx", "%rcx", "%rdx", "%rsi", "%rdi",
|
Chris@4
|
391 ; "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15"
|
Chris@4
|
392 ; );
|
Chris@4
|
393
|
Chris@4
|
394 inffas8664fnc ENDP
|
Chris@4
|
395 ;_TEXT ENDS
|
Chris@4
|
396 END
|