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}","${dir}../../perlasm");
40 open STDOUT,">$output";
42 &asm_init($ARGV[0],$0,$x86only = $ARGV[$#ARGV] eq "386");
45 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
47 &external_label("OPENSSL_ia32cap_P") if ($sse2);
51 ($a1,$a2,$a4)=("ecx","edx","ebp");
55 ($A,$B,$B30,$B31)=("mm2","mm3","mm4","mm5");
59 &function_begin_B("_mul_1x1_mmx");
62 &lea ($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 ($a2,0x7fffffff);
69 &mov (&DWP(1*4,"esp"),$a1); # a1
70 &xor ($a1,$a2); # a1^a2
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
82 &pcmpgtd($B30,$A); # broadcast 30th bit
83 &mov (&DWP(5*4,"esp"),$a1); # a1^a4
84 &xor ($a4,$a1); # a1^a2^a4
87 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
89 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
97 &movd ($R,&DWP(0,"esp",@i[0],4));
101 for($n=1;$n<9;$n++) {
102 &movd (@T[1],&DWP(0,"esp",@i[1],4));
109 push(@i,shift(@i)); push(@T,shift(@T));
111 &movd (@T[1],&DWP(0,"esp",@i[1],4));
113 &psllq (@T[1],3*$n++);
116 &movd (@T[0],&DWP(0,"esp",@i[0],4));
122 &function_end_B("_mul_1x1_mmx");
125 ($lo,$hi)=("eax","edx");
128 &function_begin_B("_mul_1x1_ialu");
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,$a1); # a1^a2^a4
148 &sar (@i[1],31); # broardcast 30th bit
150 &mov (&DWP(6*4,"esp"),$a2); # a2^a4
152 &mov (&DWP(7*4,"esp"),$a4); # a1^a2^a4
165 &mov (@i[1],0x7); # 5-byte instruction!?
169 &xor ($lo,&DWP(0,"esp",@i[0],4));
173 for($n=1;$n<9;$n++) {
174 &mov (@T[1],&DWP(0,"esp",@i[1],4));
179 &shr (@T[0],32-3*$n);
184 push(@i,shift(@i)); push(@T,shift(@T));
186 &mov (@T[1],&DWP(0,"esp",@i[1],4));
189 &mov (@i[1],&DWP(0,"esp",@i[0],4));
190 &shr (@T[0],32-3*$n); $n++;
195 &shr (@i[0],32-3*$n);
201 &function_end_B("_mul_1x1_ialu");
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");
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"));
212 &test ("eax",1<<24); # check FXSR bit
214 &test ("edx",1<<1); # check PCLMULQDQ bit
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");
224 &set_label("mmx",16);
230 &mov ($a,&wparam(1));
231 &mov ($b,&wparam(3));
232 &call ("_mul_1x1_mmx"); # a1·b1
235 &mov ($a,&wparam(2));
236 &mov ($b,&wparam(4));
237 &call ("_mul_1x1_mmx"); # a0·b0
240 &mov ($a,&wparam(1));
241 &mov ($b,&wparam(3));
242 &xor ($a,&wparam(2));
243 &xor ($b,&wparam(4));
244 &call ("_mul_1x1_mmx"); # (a0+a1)·(b0+b1)
246 &mov ($a,&wparam(0));
247 &pxor ($R,"mm6"); # (a0+a1)·(b0+b1)-a1·b1-a0·b0
257 &movq (&QWP(0,$a),$R);
259 &movq (&QWP(8,$a),$A);
262 &set_label("ialu",16);
270 &mov ($a,&wparam(1));
271 &mov ($b,&wparam(3));
272 &call ("_mul_1x1_ialu"); # a1·b1
273 &mov (&DWP(8,"esp"),$lo);
274 &mov (&DWP(12,"esp"),$hi);
276 &mov ($a,&wparam(2));
277 &mov ($b,&wparam(4));
278 &call ("_mul_1x1_ialu"); # a0·b0
279 &mov (&DWP(0,"esp"),$lo);
280 &mov (&DWP(4,"esp"),$hi);
282 &mov ($a,&wparam(1));
283 &mov ($b,&wparam(3));
284 &xor ($a,&wparam(2));
285 &xor ($b,&wparam(4));
286 &call ("_mul_1x1_ialu"); # (a0+a1)·(b0+b1)
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"));
298 &mov (&DWP(0,"ebp"),@r[0]);
300 &mov (&DWP(12,"ebp"),@r[3]);
307 &mov (&DWP(8,"ebp"),$hi);
309 &mov (&DWP(4,"ebp"),$lo);
312 &function_end_B("bn_GF2m_mul_2x2");
314 &asciz ("GF(2^m) Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");