3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> 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 # ====================================================================
10 # SHA256 block transform for x86. September 2007.
12 # Performance improvement over compiler generated code varies from
13 # 10% to 40% [see below]. Not very impressive on some ยต-archs, but
14 # it's 5 times smaller and optimizies amount of writes.
18 # Optimization including one of Pavel Semjanov's ideas resulted in
19 # ~5% improvement on P4, AMD and Sandy Bridge, and ~13% on Atom.
20 # Pavel also suggested full unroll. While his code runs ~20%/13%/6%
21 # faster on K8/Core2/Sandy Bridge, it's 9.6x larger and ~6%/18%/24%
22 # slower on P4/Atom/Pentium...
24 # Performance in clock cycles per processed byte (less is better):
26 # Pentium PIII P4 AMD K8 Core2 SB(**) Atom
29 # x86 asm 39 29 31 19 18 19(**) 31
30 # x86_64 asm(*) - - 21 16 16 18 25
32 # (*) x86_64 assembler performance is presented for reference
34 # (**) Sandy Bridge results can be improved by ~20% by replacing
35 # ror with equivalent shrd.
37 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
38 push(@INC,"${dir}","${dir}../../perlasm");
41 &asm_init($ARGV[0],"sha512-586.pl",$ARGV[$#ARGV] eq "386");
57 # *ror = sub { &shrd(@_[0],@_) };
65 &add ($T,"edi") if ($in_16_63); # T += sigma1(X[-2])
71 &mov (&DWP(4*(9+15),"esp"),$T) if ($in_16_63); # save X[0]
72 &ror ("ecx",6); # Sigma1(e)
74 &add ($T,"ecx"); # T += Sigma1(e)
75 &mov ($Eoff,$E); # modulo-scheduled
76 &xor ("esi","edi"); # Ch(e,f,g)
77 &add ($T,$Hoff); # T += h
80 &mov ($E,$Doff); # e becomes d, which is e in next iteration
82 &add ($T,"esi"); # T += Ch(e,f,g)
84 &mov ("esi",&DWP(0,$K256));
88 &mov ($Aoff,$A); # modulo-scheduled
89 &ror ("ecx",2); # Sigma0(a)
91 &xor ($A,"edi"); # a ^= b
93 &push ($A); # (b^c) in next round
94 &add ($E,$T); # d += T
95 &and ($A,&DWP(4,"esp")); # a &= (b^c)
96 &add ($T,"ecx"); # T += Sigma0(a)
97 &xor ($A,"edi"); # h = Maj(a,b,c) = Ch(a^b,c,b)
100 &add ($A,$T); # h += T
101 &mov ($T,&DWP(4*(9+15+16-1),"esp")) if ($in_16_63); # preload T
104 &function_begin("sha256_block_data_order");
105 &mov ("esi",wparam(0)); # ctx
106 &mov ("edi",wparam(1)); # inp
107 &mov ("eax",wparam(2)); # num
108 &mov ("ebx","esp"); # saved sp
110 &call (&label("pic_point")); # make it PIC!
111 &set_label("pic_point");
113 &lea ($K256,&DWP(&label("K256")."-".&label("pic_point"),$K256));
120 &mov (&DWP(0,"esp"),"esi"); # ctx
121 &mov (&DWP(4,"esp"),"edi"); # inp
122 &mov (&DWP(8,"esp"),"eax"); # inp+num*128
123 &mov (&DWP(12,"esp"),"ebx"); # saved sp
125 &set_label("loop",16);
126 # copy input block to stack reversing byte and dword order
127 for($i=0;$i<4;$i++) {
128 &mov ("eax",&DWP($i*16+0,"edi"));
129 &mov ("ebx",&DWP($i*16+4,"edi"));
130 &mov ("ecx",&DWP($i*16+8,"edi"));
132 &mov ("edx",&DWP($i*16+12,"edi"));
142 &lea ("esp",&DWP(-4*9,"esp"));# place for A,B,C,D,E,F,G,H
143 &mov (&DWP(4*(9+16)+4,"esp"),"edi");
145 # copy ctx->h[0-7] to A,B,C,D,E,F,G,H on stack
146 &mov ($A,&DWP(0,"esi"));
147 &mov ("ebx",&DWP(4,"esi"));
148 &mov ("ecx",&DWP(8,"esi"));
149 &mov ("edi",&DWP(12,"esi"));
155 &mov (&DWP(0,"esp"),"ebx"); # magic
156 &mov ($E,&DWP(16,"esi"));
157 &mov ("ebx",&DWP(20,"esi"));
158 &mov ("ecx",&DWP(24,"esi"));
159 &mov ("edi",&DWP(28,"esi"));
165 &set_label("00_15",16);
166 &mov ($T,&DWP(4*(9+15),"esp"));
170 &cmp ("esi",0xc19bf174);
171 &jne (&label("00_15"));
173 &mov ($T,&DWP(4*(9+15+16-1),"esp")); # preloaded in BODY_00_15(1)
174 &set_label("16_63",16);
176 &mov ("ecx",&DWP(4*(9+15+16-14),"esp"));
184 &xor ($T,"esi"); # T = sigma0(X[-15])
188 &add ($T,&DWP(4*(9+15+16),"esp")); # T += X[-16]
189 &xor ("edi","ecx"); # sigma1(X[-2])
191 &add ($T,&DWP(4*(9+15+16-9),"esp")); # T += X[-7]
192 # &add ($T,"edi"); # T += sigma1(X[-2])
193 # &mov (&DWP(4*(9+15),"esp"),$T); # save X[0]
197 &cmp ("esi",0xc67178f2);
198 &jne (&label("16_63"));
200 &mov ("esi",&DWP(4*(9+16+64)+0,"esp"));#ctx
205 &add ($A,&DWP(0,"esi"));
206 &add ("ebx",&DWP(4,"esi"));
207 &add ("ecx",&DWP(8,"esi"));
208 &add ("edi",&DWP(12,"esi"));
209 &mov (&DWP(0,"esi"),$A);
210 &mov (&DWP(4,"esi"),"ebx");
211 &mov (&DWP(8,"esi"),"ecx");
212 &mov (&DWP(12,"esi"),"edi");
217 &mov ("edi",&DWP(4*(9+16+64)+4,"esp"));#inp
218 &add ($E,&DWP(16,"esi"));
219 &add ("eax",&DWP(20,"esi"));
220 &add ("ebx",&DWP(24,"esi"));
221 &add ("ecx",&DWP(28,"esi"));
222 &mov (&DWP(16,"esi"),$E);
223 &mov (&DWP(20,"esi"),"eax");
224 &mov (&DWP(24,"esi"),"ebx");
225 &mov (&DWP(28,"esi"),"ecx");
227 &lea ("esp",&DWP(4*(9+16+64),"esp"));# destroy frame
228 &sub ($K256,4*64); # rewind K
230 &cmp ("edi",&DWP(8,"esp")); # are we done yet?
231 &jb (&label("loop"));
233 &mov ("esp",&DWP(12,"esp")); # restore sp
236 &set_label("K256",64); # Yes! I keep it in the code segment!
237 &data_word(0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5);
238 &data_word(0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5);
239 &data_word(0xd807aa98,0x12835b01,0x243185be,0x550c7dc3);
240 &data_word(0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174);
241 &data_word(0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc);
242 &data_word(0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da);
243 &data_word(0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7);
244 &data_word(0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967);
245 &data_word(0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13);
246 &data_word(0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85);
247 &data_word(0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3);
248 &data_word(0xd192e819,0xd6990624,0xf40e3585,0x106aa070);
249 &data_word(0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5);
250 &data_word(0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3);
251 &data_word(0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208);
252 &data_word(0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2);
253 &function_end_B("sha256_block_data_order");
254 &asciz("SHA256 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");