OLD | NEW |
1 #!/usr/local/bin/perl | 1 #!/usr/local/bin/perl |
2 | 2 |
3 push(@INC,"perlasm","../../perlasm"); | 3 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
| 4 push(@INC,"${dir}","${dir}../../perlasm"); |
4 require "x86asm.pl"; | 5 require "x86asm.pl"; |
5 | 6 |
6 &asm_init($ARGV[0],$0); | 7 &asm_init($ARGV[0],$0); |
7 | 8 |
8 $sse2=0; | 9 $sse2=0; |
9 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); } | 10 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); } |
10 | 11 |
11 &external_label("OPENSSL_ia32cap_P") if ($sse2); | 12 &external_label("OPENSSL_ia32cap_P") if ($sse2); |
12 | 13 |
13 &bn_mul_add_words("bn_mul_add_words"); | 14 &bn_mul_add_words("bn_mul_add_words"); |
14 &bn_mul_words("bn_mul_words"); | 15 &bn_mul_words("bn_mul_words"); |
15 &bn_sqr_words("bn_sqr_words"); | 16 &bn_sqr_words("bn_sqr_words"); |
16 &bn_div_words("bn_div_words"); | 17 &bn_div_words("bn_div_words"); |
17 &bn_add_words("bn_add_words"); | 18 &bn_add_words("bn_add_words"); |
18 &bn_sub_words("bn_sub_words"); | 19 &bn_sub_words("bn_sub_words"); |
19 &bn_sub_part_words("bn_sub_part_words"); | 20 &bn_sub_part_words("bn_sub_part_words"); |
20 | 21 |
21 &asm_finish(); | 22 &asm_finish(); |
22 | 23 |
23 sub bn_mul_add_words | 24 sub bn_mul_add_words |
24 { | 25 { |
25 local($name)=@_; | 26 local($name)=@_; |
26 | 27 |
27 » &function_begin($name,$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":""); | 28 » &function_begin_B($name,$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":""); |
28 | 29 |
29 » &comment(""); | 30 » $r="eax"; |
30 » $Low="eax"; | 31 » $a="edx"; |
31 » $High="edx"; | 32 » $c="ecx"; |
32 » $a="ebx"; | |
33 » $w="ebp"; | |
34 » $r="edi"; | |
35 » $c="esi"; | |
36 | |
37 » &xor($c,$c);» » # clear carry | |
38 » &mov($r,&wparam(0));» # | |
39 | |
40 » &mov("ecx",&wparam(2));»# | |
41 » &mov($a,&wparam(1));» # | |
42 | |
43 » &and("ecx",0xfffffff8);»# num / 8 | |
44 » &mov($w,&wparam(3));» # | |
45 | |
46 » &push("ecx");» » # Up the stack for a tmp variable | |
47 | |
48 » &jz(&label("maw_finish")); | |
49 | 33 |
50 if ($sse2) { | 34 if ($sse2) { |
51 &picmeup("eax","OPENSSL_ia32cap_P"); | 35 &picmeup("eax","OPENSSL_ia32cap_P"); |
52 &bt(&DWP(0,"eax"),26); | 36 &bt(&DWP(0,"eax"),26); |
53 » » &jnc(&label("maw_loop")); | 37 » » &jnc(&label("maw_non_sse2")); |
54 | 38 |
55 » » &movd("mm0",$w);» » # mm0 = w | 39 » » &mov($r,&wparam(0)); |
| 40 » » &mov($a,&wparam(1)); |
| 41 » » &mov($c,&wparam(2)); |
| 42 » » &movd("mm0",&wparam(3));» # mm0 = w |
56 &pxor("mm1","mm1"); # mm1 = carry_in | 43 &pxor("mm1","mm1"); # mm1 = carry_in |
57 | 44 » » &jmp(&label("maw_sse2_entry")); |
58 » » &set_label("maw_sse2_loop",0); | 45 » » |
| 46 » &set_label("maw_sse2_unrolled",16); |
59 &movd("mm3",&DWP(0,$r,"",0)); # mm3 = r[0] | 47 &movd("mm3",&DWP(0,$r,"",0)); # mm3 = r[0] |
60 &paddq("mm1","mm3"); # mm1 = carry_in + r[0] | 48 &paddq("mm1","mm3"); # mm1 = carry_in + r[0] |
61 &movd("mm2",&DWP(0,$a,"",0)); # mm2 = a[0] | 49 &movd("mm2",&DWP(0,$a,"",0)); # mm2 = a[0] |
62 &pmuludq("mm2","mm0"); # mm2 = w*a[0] | 50 &pmuludq("mm2","mm0"); # mm2 = w*a[0] |
63 &movd("mm4",&DWP(4,$a,"",0)); # mm4 = a[1] | 51 &movd("mm4",&DWP(4,$a,"",0)); # mm4 = a[1] |
64 &pmuludq("mm4","mm0"); # mm4 = w*a[1] | 52 &pmuludq("mm4","mm0"); # mm4 = w*a[1] |
65 &movd("mm6",&DWP(8,$a,"",0)); # mm6 = a[2] | 53 &movd("mm6",&DWP(8,$a,"",0)); # mm6 = a[2] |
66 &pmuludq("mm6","mm0"); # mm6 = w*a[2] | 54 &pmuludq("mm6","mm0"); # mm6 = w*a[2] |
67 &movd("mm7",&DWP(12,$a,"",0)); # mm7 = a[3] | 55 &movd("mm7",&DWP(12,$a,"",0)); # mm7 = a[3] |
68 &pmuludq("mm7","mm0"); # mm7 = w*a[3] | 56 &pmuludq("mm7","mm0"); # mm7 = w*a[3] |
(...skipping 36 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
105 &paddq("mm1","mm4"); # mm1 = carry4 + r[5] + w*a[5] | 93 &paddq("mm1","mm4"); # mm1 = carry4 + r[5] + w*a[5] |
106 &movd("mm5",&DWP(28,$r,"",0)); # mm5 = r[7] | 94 &movd("mm5",&DWP(28,$r,"",0)); # mm5 = r[7] |
107 &paddq("mm3","mm5"); # mm3 = r[7] + w*a[7] | 95 &paddq("mm3","mm5"); # mm3 = r[7] + w*a[7] |
108 &movd(&DWP(20,$r,"",0),"mm1"); | 96 &movd(&DWP(20,$r,"",0),"mm1"); |
109 &psrlq("mm1",32); # mm1 = carry5 | 97 &psrlq("mm1",32); # mm1 = carry5 |
110 &paddq("mm1","mm6"); # mm1 = carry5 + r[6] + w*a[6] | 98 &paddq("mm1","mm6"); # mm1 = carry5 + r[6] + w*a[6] |
111 &movd(&DWP(24,$r,"",0),"mm1"); | 99 &movd(&DWP(24,$r,"",0),"mm1"); |
112 &psrlq("mm1",32); # mm1 = carry6 | 100 &psrlq("mm1",32); # mm1 = carry6 |
113 &paddq("mm1","mm3"); # mm1 = carry6 + r[7] + w*a[7] | 101 &paddq("mm1","mm3"); # mm1 = carry6 + r[7] + w*a[7] |
114 &movd(&DWP(28,$r,"",0),"mm1"); | 102 &movd(&DWP(28,$r,"",0),"mm1"); |
115 » » &add($r,32); | 103 » » &lea($r,&DWP(32,$r)); |
116 &psrlq("mm1",32); # mm1 = carry_out | 104 &psrlq("mm1",32); # mm1 = carry_out |
117 | 105 |
118 » » &sub("ecx",8); | 106 » » &sub($c,8); |
| 107 » » &jz(&label("maw_sse2_exit")); |
| 108 » &set_label("maw_sse2_entry"); |
| 109 » » &test($c,0xfffffff8); |
| 110 » » &jnz(&label("maw_sse2_unrolled")); |
| 111 |
| 112 » &set_label("maw_sse2_loop",4); |
| 113 » » &movd("mm2",&DWP(0,$a));» # mm2 = a[i] |
| 114 » » &movd("mm3",&DWP(0,$r));» # mm3 = r[i] |
| 115 » » &pmuludq("mm2","mm0");» » # a[i] *= w |
| 116 » » &lea($a,&DWP(4,$a)); |
| 117 » » &paddq("mm1","mm3");» » # carry += r[i] |
| 118 » » &paddq("mm1","mm2");» » # carry += a[i]*w |
| 119 » » &movd(&DWP(0,$r),"mm1");» # r[i] = carry_low |
| 120 » » &sub($c,1); |
| 121 » » &psrlq("mm1",32);» » # carry = carry_high |
| 122 » » &lea($r,&DWP(4,$r)); |
119 &jnz(&label("maw_sse2_loop")); | 123 &jnz(&label("maw_sse2_loop")); |
| 124 &set_label("maw_sse2_exit"); |
| 125 &movd("eax","mm1"); # c = carry_out |
| 126 &emms(); |
| 127 &ret(); |
120 | 128 |
121 » » &movd($c,"mm1");» » # c = carry_out | 129 » &set_label("maw_non_sse2",16); |
122 » » &emms(); | |
123 | |
124 » » &jmp(&label("maw_finish")); | |
125 } | 130 } |
126 | 131 |
127 » &set_label("maw_loop",0); | 132 » # function_begin prologue |
| 133 » &push("ebp"); |
| 134 » &push("ebx"); |
| 135 » &push("esi"); |
| 136 » &push("edi"); |
128 | 137 |
129 » &mov(&swtmp(0),"ecx");» # | 138 » &comment(""); |
| 139 » $Low="eax"; |
| 140 » $High="edx"; |
| 141 » $a="ebx"; |
| 142 » $w="ebp"; |
| 143 » $r="edi"; |
| 144 » $c="esi"; |
| 145 |
| 146 » &xor($c,$c);» » # clear carry |
| 147 » &mov($r,&wparam(0));» # |
| 148 |
| 149 » &mov("ecx",&wparam(2));»# |
| 150 » &mov($a,&wparam(1));» # |
| 151 |
| 152 » &and("ecx",0xfffffff8);»# num / 8 |
| 153 » &mov($w,&wparam(3));» # |
| 154 |
| 155 » &push("ecx");» » # Up the stack for a tmp variable |
| 156 |
| 157 » &jz(&label("maw_finish")); |
| 158 |
| 159 » &set_label("maw_loop",16); |
130 | 160 |
131 for ($i=0; $i<32; $i+=4) | 161 for ($i=0; $i<32; $i+=4) |
132 { | 162 { |
133 &comment("Round $i"); | 163 &comment("Round $i"); |
134 | 164 |
135 » » &mov("eax",&DWP($i,$a,"",0)); »# *a | 165 » » &mov("eax",&DWP($i,$a)); » # *a |
136 &mul($w); # *a * w | 166 &mul($w); # *a * w |
137 » » &add("eax",$c);»» # L(t)+= *r | 167 » » &add("eax",$c);»» » # L(t)+= c |
138 » » &mov($c,&DWP($i,$r,"",0));» # L(t)+= *r | |
139 &adc("edx",0); # H(t)+=carry | 168 &adc("edx",0); # H(t)+=carry |
140 » » &add("eax",$c);» » # L(t)+=c | 169 » » &add("eax",&DWP($i,$r));» # L(t)+= *r |
141 &adc("edx",0); # H(t)+=carry | 170 &adc("edx",0); # H(t)+=carry |
142 » » &mov(&DWP($i,$r,"",0),"eax");» # *r= L(t); | 171 » » &mov(&DWP($i,$r),"eax");» # *r= L(t); |
143 &mov($c,"edx"); # c= H(t); | 172 &mov($c,"edx"); # c= H(t); |
144 } | 173 } |
145 | 174 |
146 &comment(""); | 175 &comment(""); |
147 &mov("ecx",&swtmp(0)); # | |
148 &add($a,32); | |
149 &add($r,32); | |
150 &sub("ecx",8); | 176 &sub("ecx",8); |
| 177 &lea($a,&DWP(32,$a)); |
| 178 &lea($r,&DWP(32,$r)); |
151 &jnz(&label("maw_loop")); | 179 &jnz(&label("maw_loop")); |
152 | 180 |
153 &set_label("maw_finish",0); | 181 &set_label("maw_finish",0); |
154 &mov("ecx",&wparam(2)); # get num | 182 &mov("ecx",&wparam(2)); # get num |
155 &and("ecx",7); | 183 &and("ecx",7); |
156 &jnz(&label("maw_finish2")); # helps branch prediction | 184 &jnz(&label("maw_finish2")); # helps branch prediction |
157 &jmp(&label("maw_end")); | 185 &jmp(&label("maw_end")); |
158 | 186 |
159 &set_label("maw_finish2",1); | 187 &set_label("maw_finish2",1); |
160 for ($i=0; $i<7; $i++) | 188 for ($i=0; $i<7; $i++) |
161 { | 189 { |
162 &comment("Tail Round $i"); | 190 &comment("Tail Round $i"); |
163 » » &mov("eax",&DWP($i*4,$a,"",0));# *a | 191 » » &mov("eax",&DWP($i*4,$a));» # *a |
164 &mul($w); # *a * w | 192 &mul($w); # *a * w |
165 &add("eax",$c); # L(t)+=c | 193 &add("eax",$c); # L(t)+=c |
166 &mov($c,&DWP($i*4,$r,"",0)); # L(t)+= *r | |
167 &adc("edx",0); # H(t)+=carry | 194 &adc("edx",0); # H(t)+=carry |
168 » » &add("eax",$c); | 195 » » &add("eax",&DWP($i*4,$r));» # L(t)+= *r |
169 &adc("edx",0); # H(t)+=carry | 196 &adc("edx",0); # H(t)+=carry |
170 &dec("ecx") if ($i != 7-1); | 197 &dec("ecx") if ($i != 7-1); |
171 » » &mov(&DWP($i*4,$r,"",0),"eax");»# *r= L(t); | 198 » » &mov(&DWP($i*4,$r),"eax");» # *r= L(t); |
172 » » &mov($c,"edx");» » » # c= H(t); | 199 » » &mov($c,"edx");» » # c= H(t); |
173 &jz(&label("maw_end")) if ($i != 7-1); | 200 &jz(&label("maw_end")) if ($i != 7-1); |
174 } | 201 } |
175 &set_label("maw_end",0); | 202 &set_label("maw_end",0); |
176 &mov("eax",$c); | 203 &mov("eax",$c); |
177 | 204 |
178 &pop("ecx"); # clear variable from | 205 &pop("ecx"); # clear variable from |
179 | 206 |
180 &function_end($name); | 207 &function_end($name); |
181 } | 208 } |
182 | 209 |
183 sub bn_mul_words | 210 sub bn_mul_words |
184 { | 211 { |
185 local($name)=@_; | 212 local($name)=@_; |
186 | 213 |
187 » &function_begin($name,""); | 214 » &function_begin_B($name,$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":""); |
| 215 |
| 216 » $r="eax"; |
| 217 » $a="edx"; |
| 218 » $c="ecx"; |
| 219 |
| 220 » if ($sse2) { |
| 221 » » &picmeup("eax","OPENSSL_ia32cap_P"); |
| 222 » » &bt(&DWP(0,"eax"),26); |
| 223 » » &jnc(&label("mw_non_sse2")); |
| 224 |
| 225 » » &mov($r,&wparam(0)); |
| 226 » » &mov($a,&wparam(1)); |
| 227 » » &mov($c,&wparam(2)); |
| 228 » » &movd("mm0",&wparam(3));» # mm0 = w |
| 229 » » &pxor("mm1","mm1");» » # mm1 = carry = 0 |
| 230 |
| 231 » &set_label("mw_sse2_loop",16); |
| 232 » » &movd("mm2",&DWP(0,$a));» # mm2 = a[i] |
| 233 » » &pmuludq("mm2","mm0");» » # a[i] *= w |
| 234 » » &lea($a,&DWP(4,$a)); |
| 235 » » &paddq("mm1","mm2");» » # carry += a[i]*w |
| 236 » » &movd(&DWP(0,$r),"mm1");» # r[i] = carry_low |
| 237 » » &sub($c,1); |
| 238 » » &psrlq("mm1",32);» » # carry = carry_high |
| 239 » » &lea($r,&DWP(4,$r)); |
| 240 » » &jnz(&label("mw_sse2_loop")); |
| 241 |
| 242 » » &movd("eax","mm1");» » # return carry |
| 243 » » &emms(); |
| 244 » » &ret(); |
| 245 » &set_label("mw_non_sse2",16); |
| 246 » } |
| 247 |
| 248 » # function_begin prologue |
| 249 » &push("ebp"); |
| 250 » &push("ebx"); |
| 251 » &push("esi"); |
| 252 » &push("edi"); |
188 | 253 |
189 &comment(""); | 254 &comment(""); |
190 $Low="eax"; | 255 $Low="eax"; |
191 $High="edx"; | 256 $High="edx"; |
192 $a="ebx"; | 257 $a="ebx"; |
193 $w="ecx"; | 258 $w="ecx"; |
194 $r="edi"; | 259 $r="edi"; |
195 $c="esi"; | 260 $c="esi"; |
196 $num="ebp"; | 261 $num="ebp"; |
197 | 262 |
(...skipping 52 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
250 &set_label("mw_end",0); | 315 &set_label("mw_end",0); |
251 &mov("eax",$c); | 316 &mov("eax",$c); |
252 | 317 |
253 &function_end($name); | 318 &function_end($name); |
254 } | 319 } |
255 | 320 |
256 sub bn_sqr_words | 321 sub bn_sqr_words |
257 { | 322 { |
258 local($name)=@_; | 323 local($name)=@_; |
259 | 324 |
260 » &function_begin($name,""); | 325 » &function_begin_B($name,$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":""); |
| 326 |
| 327 » $r="eax"; |
| 328 » $a="edx"; |
| 329 » $c="ecx"; |
| 330 |
| 331 » if ($sse2) { |
| 332 » » &picmeup("eax","OPENSSL_ia32cap_P"); |
| 333 » » &bt(&DWP(0,"eax"),26); |
| 334 » » &jnc(&label("sqr_non_sse2")); |
| 335 |
| 336 » » &mov($r,&wparam(0)); |
| 337 » » &mov($a,&wparam(1)); |
| 338 » » &mov($c,&wparam(2)); |
| 339 |
| 340 » &set_label("sqr_sse2_loop",16); |
| 341 » » &movd("mm0",&DWP(0,$a));» # mm0 = a[i] |
| 342 » » &pmuludq("mm0","mm0");» » # a[i] *= a[i] |
| 343 » » &lea($a,&DWP(4,$a));» » # a++ |
| 344 » » &movq(&QWP(0,$r),"mm0");» # r[i] = a[i]*a[i] |
| 345 » » &sub($c,1); |
| 346 » » &lea($r,&DWP(8,$r));» » # r += 2 |
| 347 » » &jnz(&label("sqr_sse2_loop")); |
| 348 |
| 349 » » &emms(); |
| 350 » » &ret(); |
| 351 » &set_label("sqr_non_sse2",16); |
| 352 » } |
| 353 |
| 354 » # function_begin prologue |
| 355 » &push("ebp"); |
| 356 » &push("ebx"); |
| 357 » &push("esi"); |
| 358 » &push("edi"); |
261 | 359 |
262 &comment(""); | 360 &comment(""); |
263 $r="esi"; | 361 $r="esi"; |
264 $a="edi"; | 362 $a="edi"; |
265 $num="ebx"; | 363 $num="ebx"; |
266 | 364 |
267 &mov($r,&wparam(0)); # | 365 &mov($r,&wparam(0)); # |
268 &mov($a,&wparam(1)); # | 366 &mov($a,&wparam(1)); # |
269 &mov($num,&wparam(2)); # | 367 &mov($num,&wparam(2)); # |
270 | 368 |
(...skipping 35 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
306 } | 404 } |
307 &set_label("sw_end",0); | 405 &set_label("sw_end",0); |
308 | 406 |
309 &function_end($name); | 407 &function_end($name); |
310 } | 408 } |
311 | 409 |
312 sub bn_div_words | 410 sub bn_div_words |
313 { | 411 { |
314 local($name)=@_; | 412 local($name)=@_; |
315 | 413 |
316 » &function_begin($name,""); | 414 » &function_begin_B($name,""); |
317 &mov("edx",&wparam(0)); # | 415 &mov("edx",&wparam(0)); # |
318 &mov("eax",&wparam(1)); # | 416 &mov("eax",&wparam(1)); # |
319 » &mov("ebx",&wparam(2));»# | 417 » &mov("ecx",&wparam(2));»# |
320 » &div("ebx"); | 418 » &div("ecx"); |
321 » &function_end($name); | 419 » &ret(); |
| 420 » &function_end_B($name); |
322 } | 421 } |
323 | 422 |
324 sub bn_add_words | 423 sub bn_add_words |
325 { | 424 { |
326 local($name)=@_; | 425 local($name)=@_; |
327 | 426 |
328 &function_begin($name,""); | 427 &function_begin($name,""); |
329 | 428 |
330 &comment(""); | 429 &comment(""); |
331 $a="esi"; | 430 $a="esi"; |
(...skipping 334 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
666 &set_label("pw_nc_end",0); | 765 &set_label("pw_nc_end",0); |
667 &mov($c,0); | 766 &mov($c,0); |
668 | 767 |
669 &set_label("pw_end",0); | 768 &set_label("pw_end",0); |
670 | 769 |
671 # &mov("eax",$c); # $c is "eax" | 770 # &mov("eax",$c); # $c is "eax" |
672 | 771 |
673 &function_end($name); | 772 &function_end($name); |
674 } | 773 } |
675 | 774 |
OLD | NEW |