]> git.ipfire.org Git - thirdparty/openssl.git/blame - crypto/bn/asm/x86-gf2m.pl
projects / thirdparty / openssl.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Doc nits cleanup, round 2
[thirdparty/openssl.git] / crypto / bn / asm / x86-gf2m.pl
CommitLineData
58cc21fd
AP		 1#!/usr/bin/env perl
2#
3# ====================================================================
4# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9#
10# May 2011
11#
12# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
13# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
14# the time being... Except that it has three code paths: pure integer
15# code suitable for any x86 CPU, MMX code suitable for PIII and later
16# and PCLMULQDQ suitable for Westmere and later. Improvement varies
053fa39a 17# from one benchmark and µ-arch to another. Below are interval values
58cc21fd
AP		 18# for 163- and 571-bit ECDH benchmarks relative to compiler-generated
19# code:
20#
21# PIII 16%-30%
22# P4 12%-12%
23# Opteron 18%-40%
24# Core2 19%-44%
25# Atom 38%-64%
26# Westmere 53%-121%(PCLMULQDQ)/20%-32%(MMX)
27# Sandy Bridge 72%-127%(PCLMULQDQ)/27%-23%(MMX)
28#
29# Note that above improvement coefficients are not coefficients for
30# bn_GF2m_mul_2x2 itself. For example 120% ECDH improvement is result
31# of bn_GF2m_mul_2x2 being >4x faster. As it gets faster, benchmark
32# is more and more dominated by other subroutines, most notably by
33# BN_GF2m_mod[_mul]_arr...
34
35$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
36push(@INC,"${dir}","${dir}../../perlasm");
37require "x86asm.pl";
38
6bd7a4d9
RL		 39$output = pop;
40open STDOUT,">$output";
41
58cc21fd
AP		 42&asm_init($ARGV[0],$0,$x86only = $ARGV[$#ARGV] eq "386");
43
44$sse2=0;
45for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
46
47&external_label("OPENSSL_ia32cap_P") if ($sse2);
48
49$a="eax";
50$b="ebx";
51($a1,$a2,$a4)=("ecx","edx","ebp");
52
53$R="mm0";
54@T=("mm1","mm2");
55($A,$B,$B30,$B31)=("mm2","mm3","mm4","mm5");
56@i=("esi","edi");
57
58 if (!$x86only) {
59&function_begin_B("_mul_1x1_mmx");
60 &sub ("esp",32+4);
61 &mov ($a1,$a);
62 &lea ($a2,&DWP(0,$a,$a));
63 &and ($a1,0x3fffffff);
64 &lea ($a4,&DWP(0,$a2,$a2));
65 &mov (&DWP(0*4,"esp"),0);
66 &and ($a2,0x7fffffff);
67 &movd ($A,$a);
68 &movd ($B,$b);
69 &mov (&DWP(1*4,"esp"),$a1); # a1
70 &xor ($a1,$a2); # a1^a2
71 &pxor ($B31,$B31);
72 &pxor ($B30,$B30);
73 &mov (&DWP(2*4,"esp"),$a2); # a2
74 &xor ($a2,$a4); # a2^a4
75 &mov (&DWP(3*4,"esp"),$a1); # a1^a2
76 &pcmpgtd($B31,$A); # broadcast 31st bit
77 &paddd ($A,$A); # $A<<=1
78 &xor ($a1,$a2); # a1^a4=a1^a2^a2^a4
79 &mov (&DWP(4*4,"esp"),$a4); # a4
80 &xor ($a4,$a2); # a2=a4^a2^a4
81 &pand ($B31,$B);
82 &pcmpgtd($B30,$A); # broadcast 30th bit
83 &mov (&DWP(5*4,"esp"),$a1); # a1^a4
84 &xor ($a4,$a1); # a1^a2^a4
85 &psllq ($B31,31);
86 &pand ($B30,$B);
87 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
88 &mov (@i[0],0x7);
89 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
90 &mov ($a4,@i[0]);
91 &and (@i[0],$b);
92 &shr ($b,3);
93 &mov (@i[1],$a4);
94 &psllq ($B30,30);
95 &and (@i[1],$b);
96 &shr ($b,3);
97 &movd ($R,&DWP(0,"esp",@i[0],4));
98 &mov (@i[0],$a4);
99 &and (@i[0],$b);
100 &shr ($b,3);
101 for($n=1;$n<9;$n++) {
102 &movd (@T[1],&DWP(0,"esp",@i[1],4));
103 &mov (@i[1],$a4);
104 &psllq (@T[1],3*$n);
105 &and (@i[1],$b);
106 &shr ($b,3);
107 &pxor ($R,@T[1]);
108
109 push(@i,shift(@i)); push(@T,shift(@T));
110 }
111 &movd (@T[1],&DWP(0,"esp",@i[1],4));
112 &pxor ($R,$B30);
113 &psllq (@T[1],3*$n++);
114 &pxor ($R,@T[1]);
115
116 &movd (@T[0],&DWP(0,"esp",@i[0],4));
117 &pxor ($R,$B31);
118 &psllq (@T[0],3*$n);
119 &add ("esp",32+4);
120 &pxor ($R,@T[0]);
121 &ret ();
122&function_end_B("_mul_1x1_mmx");
123 }
124
125($lo,$hi)=("eax","edx");
126@T=("ecx","ebp");
127
128&function_begin_B("_mul_1x1_ialu");
129 &sub ("esp",32+4);
130 &mov ($a1,$a);
131 &lea ($a2,&DWP(0,$a,$a));
132 &lea ($a4,&DWP(0,"",$a,4));
133 &and ($a1,0x3fffffff);
134 &lea (@i[1],&DWP(0,$lo,$lo));
135 &sar ($lo,31); # broadcast 31st bit
136 &mov (&DWP(0*4,"esp"),0);
137 &and ($a2,0x7fffffff);
138 &mov (&DWP(1*4,"esp"),$a1); # a1
139 &xor ($a1,$a2); # a1^a2
140 &mov (&DWP(2*4,"esp"),$a2); # a2
141 &xor ($a2,$a4); # a2^a4
142 &mov (&DWP(3*4,"esp"),$a1); # a1^a2
143 &xor ($a1,$a2); # a1^a4=a1^a2^a2^a4
144 &mov (&DWP(4*4,"esp"),$a4); # a4
145 &xor ($a4,$a2); # a2=a4^a2^a4
146 &mov (&DWP(5*4,"esp"),$a1); # a1^a4
147 &xor ($a4,$a1); # a1^a2^a4
148 &sar (@i[1],31); # broardcast 30th bit
149 &and ($lo,$b);
150 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
151 &and (@i[1],$b);
152 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
153 &mov ($hi,$lo);
154 &shl ($lo,31);
155 &mov (@T[0],@i[1]);
156 &shr ($hi,1);
157
158 &mov (@i[0],0x7);
159 &shl (@i[1],30);
160 &and (@i[0],$b);
161 &shr (@T[0],2);
162 &xor ($lo,@i[1]);
163
164 &shr ($b,3);
165 &mov (@i[1],0x7); # 5-byte instruction!?
166 &and (@i[1],$b);
167 &shr ($b,3);
168 &xor ($hi,@T[0]);
169 &xor ($lo,&DWP(0,"esp",@i[0],4));
170 &mov (@i[0],0x7);
171 &and (@i[0],$b);
172 &shr ($b,3);
173 for($n=1;$n<9;$n++) {
174 &mov (@T[1],&DWP(0,"esp",@i[1],4));
175 &mov (@i[1],0x7);
176 &mov (@T[0],@T[1]);
177 &shl (@T[1],3*$n);
178 &and (@i[1],$b);
179 &shr (@T[0],32-3*$n);
180 &xor ($lo,@T[1]);
181 &shr ($b,3);
182 &xor ($hi,@T[0]);
183
184 push(@i,shift(@i)); push(@T,shift(@T));
185 }
186 &mov (@T[1],&DWP(0,"esp",@i[1],4));
187 &mov (@T[0],@T[1]);
188 &shl (@T[1],3*$n);
189 &mov (@i[1],&DWP(0,"esp",@i[0],4));
190 &shr (@T[0],32-3*$n); $n++;
191 &mov (@i[0],@i[1]);
192 &xor ($lo,@T[1]);
193 &shl (@i[1],3*$n);
194 &xor ($hi,@T[0]);
195 &shr (@i[0],32-3*$n);
196 &xor ($lo,@i[1]);
197 &xor ($hi,@i[0]);
198
199 &add ("esp",32+4);
200 &ret ();
201&function_end_B("_mul_1x1_ialu");
202
58cc21fd
AP		 203# void bn_GF2m_mul_2x2(BN_ULONG *r, BN_ULONG a1, BN_ULONG a0, BN_ULONG b1, BN_ULONG b0);
204&function_begin_B("bn_GF2m_mul_2x2");
205if (!$x86only) {
206 &picmeup("edx","OPENSSL_ia32cap_P");
207 &mov ("eax",&DWP(0,"edx"));
208 &mov ("edx",&DWP(4,"edx"));
209 &test ("eax",1<<23); # check MMX bit
210 &jz (&label("ialu"));
211if ($sse2) {
212 &test ("eax",1<<24); # check FXSR bit
213 &jz (&label("mmx"));
214 &test ("edx",1<<1); # check PCLMULQDQ bit
215 &jz (&label("mmx"));
216
217 &movups ("xmm0",&QWP(8,"esp"));
218 &shufps ("xmm0","xmm0",0b10110001);
219 &pclmulqdq ("xmm0","xmm0",1);
220 &mov ("eax",&DWP(4,"esp"));
221 &movups (&QWP(0,"eax"),"xmm0");
222 &ret ();
223
224&set_label("mmx",16);
225}
226 &push ("ebp");
227 &push ("ebx");
228 &push ("esi");
229 &push ("edi");
230 &mov ($a,&wparam(1));
231 &mov ($b,&wparam(3));
053fa39a 232 &call ("_mul_1x1_mmx"); # a1·b1
58cc21fd
AP		 233 &movq ("mm7",$R);
234
235 &mov ($a,&wparam(2));
236 &mov ($b,&wparam(4));
053fa39a 237 &call ("_mul_1x1_mmx"); # a0·b0
58cc21fd
AP		 238 &movq ("mm6",$R);
239
240 &mov ($a,&wparam(1));
241 &mov ($b,&wparam(3));
242 &xor ($a,&wparam(2));
243 &xor ($b,&wparam(4));
053fa39a 244 &call ("_mul_1x1_mmx"); # (a0+a1)·(b0+b1)
58cc21fd
AP		 245 &pxor ($R,"mm7");
246 &mov ($a,&wparam(0));
053fa39a 247 &pxor ($R,"mm6"); # (a0+a1)·(b0+b1)-a1·b1-a0·b0
58cc21fd
AP		 248
249 &movq ($A,$R);
250 &psllq ($R,32);
251 &pop ("edi");
252 &psrlq ($A,32);
253 &pop ("esi");
254 &pxor ($R,"mm6");
255 &pop ("ebx");
256 &pxor ($A,"mm7");
257 &movq (&QWP(0,$a),$R);
258 &pop ("ebp");
259 &movq (&QWP(8,$a),$A);
260 &emms ();
261 &ret ();
262&set_label("ialu",16);
263}
264 &push ("ebp");
265 &push ("ebx");
266 &push ("esi");
267 &push ("edi");
268 &stack_push(4+1);
269
270 &mov ($a,&wparam(1));
271 &mov ($b,&wparam(3));
053fa39a 272 &call ("_mul_1x1_ialu"); # a1·b1
58cc21fd
AP		 273 &mov (&DWP(8,"esp"),$lo);
274 &mov (&DWP(12,"esp"),$hi);
275
276 &mov ($a,&wparam(2));
277 &mov ($b,&wparam(4));
053fa39a 278 &call ("_mul_1x1_ialu"); # a0·b0
58cc21fd
AP		 279 &mov (&DWP(0,"esp"),$lo);
280 &mov (&DWP(4,"esp"),$hi);
281
282 &mov ($a,&wparam(1));
283 &mov ($b,&wparam(3));
284 &xor ($a,&wparam(2));
285 &xor ($b,&wparam(4));
053fa39a 286 &call ("_mul_1x1_ialu"); # (a0+a1)·(b0+b1)
58cc21fd
AP		 287
288 &mov ("ebp",&wparam(0));
289 @r=("ebx","ecx","edi","esi");
290 &mov (@r[0],&DWP(0,"esp"));
291 &mov (@r[1],&DWP(4,"esp"));
292 &mov (@r[2],&DWP(8,"esp"));
293 &mov (@r[3],&DWP(12,"esp"));
294
295 &xor ($lo,$hi);
296 &xor ($hi,@r[1]);
297 &xor ($lo,@r[0]);
298 &mov (&DWP(0,"ebp"),@r[0]);
299 &xor ($hi,@r[2]);
300 &mov (&DWP(12,"ebp"),@r[3]);
301 &xor ($lo,@r[3]);
302 &stack_pop(4+1);
303 &xor ($hi,@r[3]);
304 &pop ("edi");
305 &xor ($lo,$hi);
306 &pop ("esi");
307 &mov (&DWP(8,"ebp"),$hi);
308 &pop ("ebx");
309 &mov (&DWP(4,"ebp"),$lo);
310 &pop ("ebp");
311 &ret ();
312&function_end_B("bn_GF2m_mul_2x2");
313
2b9a8ca1 314&asciz ("GF(2^m) Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");
58cc21fd
AP		 315
316&asm_finish();
6bd7a4d9
RL		 317
318close STDOUT;
A mirror of the official openssl repository