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 # ====================================================================
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
17 # from one benchmark and µ-arch to another. Below are interval values
18 # for 163- and 571-bit ECDH benchmarks relative to compiler-generated
26 # Westmere 53%-121%(PCLMULQDQ)/20%-32%(MMX)
27 # Sandy Bridge 72%-127%(PCLMULQDQ)/27%-23%(MMX)
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...
35 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
36 push(@INC,"${dir}","${dir}../../perlasm");
39 &asm_init($ARGV[0],$0,$x86only = $ARGV[$#ARGV] eq "386");
42 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
44 &external_label("OPENSSL_ia32cap_P") if ($sse2);
48 ($a1,$a2,$a4)=("ecx","edx","ebp");
52 ($A,$B,$B30,$B31)=("mm2","mm3","mm4","mm5");
56 &function_begin_B("_mul_1x1_mmx");
59 &lea ($a2,&DWP(0,$a,$a));
60 &and ($a1,0x3fffffff);
61 &lea ($a4,&DWP(0,$a2,$a2));
62 &mov (&DWP(0*4,"esp"),0);
63 &and ($a2,0x7fffffff);
66 &mov (&DWP(1*4,"esp"),$a1); # a1
67 &xor ($a1,$a2); # a1^a2
70 &mov (&DWP(2*4,"esp"),$a2); # a2
71 &xor ($a2,$a4); # a2^a4
72 &mov (&DWP(3*4,"esp"),$a1); # a1^a2
73 &pcmpgtd($B31,$A); # broadcast 31st bit
74 &paddd ($A,$A); # $A<<=1
75 &xor ($a1,$a2); # a1^a4=a1^a2^a2^a4
76 &mov (&DWP(4*4,"esp"),$a4); # a4
77 &xor ($a4,$a2); # a2=a4^a2^a4
79 &pcmpgtd($B30,$A); # broadcast 30th bit
80 &mov (&DWP(5*4,"esp"),$a1); # a1^a4
81 &xor ($a4,$a1); # a1^a2^a4
84 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
86 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
94 &movd ($R,&DWP(0,"esp",@i[0],4));
99 &movd (@T[1],&DWP(0,"esp",@i[1],4));
106 push(@i,shift(@i)); push(@T,shift(@T));
108 &movd (@T[1],&DWP(0,"esp",@i[1],4));
110 &psllq (@T[1],3*$n++);
113 &movd (@T[0],&DWP(0,"esp",@i[0],4));
119 &function_end_B("_mul_1x1_mmx");
122 ($lo,$hi)=("eax","edx");
125 &function_begin_B("_mul_1x1_ialu");
128 &lea ($a2,&DWP(0,$a,$a));
129 &lea ($a4,&DWP(0,"",$a,4));
130 &and ($a1,0x3fffffff);
131 &lea (@i[1],&DWP(0,$lo,$lo));
132 &sar ($lo,31); # broadcast 31st bit
133 &mov (&DWP(0*4,"esp"),0);
134 &and ($a2,0x7fffffff);
135 &mov (&DWP(1*4,"esp"),$a1); # a1
136 &xor ($a1,$a2); # a1^a2
137 &mov (&DWP(2*4,"esp"),$a2); # a2
138 &xor ($a2,$a4); # a2^a4
139 &mov (&DWP(3*4,"esp"),$a1); # a1^a2
140 &xor ($a1,$a2); # a1^a4=a1^a2^a2^a4
141 &mov (&DWP(4*4,"esp"),$a4); # a4
142 &xor ($a4,$a2); # a2=a4^a2^a4
143 &mov (&DWP(5*4,"esp"),$a1); # a1^a4
144 &xor ($a4,$a1); # a1^a2^a4
145 &sar (@i[1],31); # broardcast 30th bit
147 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
149 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
162 &mov (@i[1],0x7); # 5-byte instruction!?
166 &xor ($lo,&DWP(0,"esp",@i[0],4));
170 for($n=1;$n<9;$n++) {
171 &mov (@T[1],&DWP(0,"esp",@i[1],4));
176 &shr (@T[0],32-3*$n);
181 push(@i,shift(@i)); push(@T,shift(@T));
183 &mov (@T[1],&DWP(0,"esp",@i[1],4));
186 &mov (@i[1],&DWP(0,"esp",@i[0],4));
187 &shr (@T[0],32-3*$n); $n++;
192 &shr (@i[0],32-3*$n);
198 &function_end_B("_mul_1x1_ialu");
200 # void bn_GF2m_mul_2x2(BN_ULONG *r, BN_ULONG a1, BN_ULONG a0, BN_ULONG b1, BN_ULONG b0);
201 &function_begin_B("bn_GF2m_mul_2x2");
203 &picmeup("edx","OPENSSL_ia32cap_P");
204 &mov ("eax",&DWP(0,"edx"));
205 &mov ("edx",&DWP(4,"edx"));
206 &test ("eax",1<<23); # check MMX bit
207 &jz (&label("ialu"));
209 &test ("eax",1<<24); # check FXSR bit
211 &test ("edx",1<<1); # check PCLMULQDQ bit
214 &movups ("xmm0",&QWP(8,"esp"));
215 &shufps ("xmm0","xmm0",0b10110001);
216 &pclmulqdq ("xmm0","xmm0",1);
217 &mov ("eax",&DWP(4,"esp"));
218 &movups (&QWP(0,"eax"),"xmm0");
221 &set_label("mmx",16);
227 &mov ($a,&wparam(1));
228 &mov ($b,&wparam(3));
229 &call ("_mul_1x1_mmx"); # a1·b1
232 &mov ($a,&wparam(2));
233 &mov ($b,&wparam(4));
234 &call ("_mul_1x1_mmx"); # a0·b0
237 &mov ($a,&wparam(1));
238 &mov ($b,&wparam(3));
239 &xor ($a,&wparam(2));
240 &xor ($b,&wparam(4));
241 &call ("_mul_1x1_mmx"); # (a0+a1)·(b0+b1)
243 &mov ($a,&wparam(0));
244 &pxor ($R,"mm6"); # (a0+a1)·(b0+b1)-a1·b1-a0·b0
254 &movq (&QWP(0,$a),$R);
256 &movq (&QWP(8,$a),$A);
259 &set_label("ialu",16);
267 &mov ($a,&wparam(1));
268 &mov ($b,&wparam(3));
269 &call ("_mul_1x1_ialu"); # a1·b1
270 &mov (&DWP(8,"esp"),$lo);
271 &mov (&DWP(12,"esp"),$hi);
273 &mov ($a,&wparam(2));
274 &mov ($b,&wparam(4));
275 &call ("_mul_1x1_ialu"); # a0·b0
276 &mov (&DWP(0,"esp"),$lo);
277 &mov (&DWP(4,"esp"),$hi);
279 &mov ($a,&wparam(1));
280 &mov ($b,&wparam(3));
281 &xor ($a,&wparam(2));
282 &xor ($b,&wparam(4));
283 &call ("_mul_1x1_ialu"); # (a0+a1)·(b0+b1)
285 &mov ("ebp",&wparam(0));
286 @r=("ebx","ecx","edi","esi");
287 &mov (@r[0],&DWP(0,"esp"));
288 &mov (@r[1],&DWP(4,"esp"));
289 &mov (@r[2],&DWP(8,"esp"));
290 &mov (@r[3],&DWP(12,"esp"));
295 &mov (&DWP(0,"ebp"),@r[0]);
297 &mov (&DWP(12,"ebp"),@r[3]);
304 &mov (&DWP(8,"ebp"),$hi);
306 &mov (&DWP(4,"ebp"),$lo);
309 &function_end_B("bn_GF2m_mul_2x2");
311 &asciz ("GF(2^m) Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");