Commentary section update in sha512-x86_64.pl.
[openssl.git] / crypto / sha / asm / sha512-x86_64.pl
1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
8 #
9 # sha256/512_block procedure for x86_64.
10 #
11 # 40% improvement over compiler-generated code on Opteron. On EM64T
12 # sha256 was observed to run >80% faster and sha512 - >40%. No magical
13 # tricks, just straight implementation... I really wonder why gcc
14 # [being armed with inline assembler] fails to generate as fast code.
15 # The only thing which is cool about this module is that it's very
16 # same instruction sequence used for both SHA-256 and SHA-512. In
17 # former case the instructions operate on 32-bit operands, while in
18 # latter - on 64-bit ones. All I had to do is to get one flavor right,
19 # the other one passed the test right away:-)
20 #
21 # sha256_block runs in ~1005 cycles on Opteron, which gives you
22 # asymptotic performance of 64*1000/1005=63.7MBps times CPU clock
23 # frequency in GHz. sha512_block runs in ~1275 cycles, which results
24 # in 128*1000/1275=100MBps per GHz. Is there room for improvement?
25 # Well, if you compare it to IA-64 implementation, which maintains
26 # X[16] in register bank[!], tends to 4 instructions per CPU clock
27 # cycle and runs in 1003 cycles, 1275 is very good result for 3-way
28 # issue Opteron pipeline and X[16] maintained in memory. So that *if*
29 # there is a way to improve it, *then* the only way would be to try to
30 # offload X[16] updates to SSE unit, but that would require "deeper"
31 # loop unroll, which in turn would naturally cause size blow-up, not
32 # to mention increased complexity! And once again, only *if* it's
33 # actually possible to noticeably improve overall ILP, instruction
34 # level parallelism, on a given CPU implementation in this case.
35 #
36 # Special note on Intel EM64T. While Opteron CPU exhibits perfect
37 # perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
38 # [currently available] EM64T CPUs apparently are far from it. On the
39 # contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
40 # sha256_block:-( This is presumably because 64-bit shifts/rotates
41 # apparently are not atomic instructions, but implemented in microcode.
42
43 $output=shift;
44 open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
45
46 if ($output =~ /512/) {
47         $func="sha512_block";
48         $TABLE="K512";
49         $SZ=8;
50         @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
51                                         "%r8", "%r9", "%r10","%r11");
52         ($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
53         @Sigma0=(28,34,39);
54         @Sigma1=(14,18,41);
55         @sigma0=(1,  8, 7);
56         @sigma1=(19,61, 6);
57         $rounds=80;
58 } else {
59         $func="sha256_block";
60         $TABLE="K256";
61         $SZ=4;
62         @ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
63                                         "%r8d","%r9d","%r10d","%r11d");
64         ($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
65         @Sigma0=( 2,13,22);
66         @Sigma1=( 6,11,25);
67         @sigma0=( 7,18, 3);
68         @sigma1=(17,19,10);
69         $rounds=64;
70 }
71
72 $ctx="%rdi";    # 1st arg
73 $round="%rdi";  # zaps $ctx
74 $inp="%rsi";    # 2nd arg
75 $Tbl="%rbp";
76
77 $_ctx="16*$SZ+0*8(%rsp)";
78 $_inp="16*$SZ+1*8(%rsp)";
79 $_end="16*$SZ+2*8(%rsp)";
80 $_ord="16*$SZ+3*8(%rsp)";
81 $_rsp="16*$SZ+4*8(%rsp)";
82 $framesz="16*$SZ+5*8";
83
84
85 sub ROUND_00_15()
86 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
87
88 $code.=<<___;
89         mov     $e,$a0
90         mov     $e,$a1
91         mov     $f,$a2
92
93         ror     \$$Sigma1[0],$a0
94         ror     \$$Sigma1[1],$a1
95         xor     $g,$a2                  # f^g
96
97         xor     $a1,$a0
98         ror     \$`$Sigma1[2]-$Sigma1[1]`,$a1
99         and     $e,$a2                  # (f^g)&e
100         mov     $T1,`$SZ*($i&0xf)`(%rsp)
101
102         xor     $a1,$a0                 # Sigma1(e)
103         xor     $g,$a2                  # Ch(e,f,g)=((f^g)&e)^g
104         add     $h,$T1                  # T1+=h
105
106         mov     $a,$h
107         add     $a0,$T1                 # T1+=Sigma1(e)
108
109         add     $a2,$T1                 # T1+=Ch(e,f,g)
110         mov     $a,$a0
111         mov     $a,$a1
112
113         ror     \$$Sigma0[0],$h
114         ror     \$$Sigma0[1],$a0
115         mov     $a,$a2
116         add     ($Tbl,$round,$SZ),$T1   # T1+=K[round]
117
118         xor     $a0,$h
119         ror     \$`$Sigma0[2]-$Sigma0[1]`,$a0
120         or      $c,$a1                  # a|c
121
122         xor     $a0,$h                  # h=Sigma0(a)
123         and     $c,$a2                  # a&c
124         add     $T1,$d                  # d+=T1
125
126         and     $b,$a1                  # (a|c)&b
127         add     $T1,$h                  # h+=T1
128
129         or      $a2,$a1                 # Maj(a,b,c)=((a|c)&b)|(a&c)
130         lea     1($round),$round        # round++
131
132         add     $a1,$h                  # h+=Maj(a,b,c)
133 ___
134 }
135
136 sub ROUND_16_XX()
137 { my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
138
139 $code.=<<___;
140         mov     `$SZ*(($i+1)&0xf)`(%rsp),$a0
141         mov     `$SZ*(($i+14)&0xf)`(%rsp),$T1
142
143         mov     $a0,$a2
144
145         shr     \$$sigma0[2],$a0
146         ror     \$$sigma0[0],$a2
147
148         xor     $a2,$a0
149         ror     \$`$sigma0[1]-$sigma0[0]`,$a2
150
151         xor     $a2,$a0                 # sigma0(X[(i+1)&0xf])
152         mov     $T1,$a1
153
154         shr     \$$sigma1[2],$T1
155         ror     \$$sigma1[0],$a1
156
157         xor     $a1,$T1
158         ror     \$`$sigma1[1]-$sigma1[0]`,$a1
159
160         xor     $a1,$T1                 # sigma1(X[(i+14)&0xf])
161
162         add     $a0,$T1
163
164         add     `$SZ*(($i+9)&0xf)`(%rsp),$T1
165
166         add     `$SZ*($i&0xf)`(%rsp),$T1
167 ___
168         &ROUND_00_15(@_);
169 }
170
171 $code=<<___;
172 .text
173
174 .globl  $func
175 .type   $func,\@function,4
176 .align  16
177 $func:
178         push    %rbx
179         push    %rbp
180         push    %r12
181         push    %r13
182         push    %r14
183         push    %r15
184         mov     %rsp,%rbp               # copy %rsp
185         shl     \$4,%rdx                # num*16
186         sub     \$$framesz,%rsp
187         lea     ($inp,%rdx,$SZ),%rdx    # inp+num*16*$SZ
188         and     \$-64,%rsp              # align stack frame
189         mov     $ctx,$_ctx              # save ctx, 1st arg
190         mov     $inp,$_inp              # save inp, 2nd arh
191         mov     %rdx,$_end              # save end pointer, "3rd" arg
192         mov     %ecx,$_ord              # save host, 4th arg
193         mov     %rbp,$_rsp              # save copy of %rsp
194
195         .picmeup $Tbl
196         lea     $TABLE-.($Tbl),$Tbl
197
198         mov     $SZ*0($ctx),$A
199         mov     $SZ*1($ctx),$B
200         mov     $SZ*2($ctx),$C
201         mov     $SZ*3($ctx),$D
202         mov     $SZ*4($ctx),$E
203         mov     $SZ*5($ctx),$F
204         mov     $SZ*6($ctx),$G
205         mov     $SZ*7($ctx),$H
206         jmp     .Lloop
207
208 .align  16
209 .Lloop:
210         xor     $round,$round
211 ___
212 if ($SZ==4) {
213 $code.=<<___;
214         cmpl    \$0,$_ord
215         je      .Ldata_order
216 .align  16
217 .Lhost_order:
218 ___
219
220         for($i=0;$i<16;$i++) {
221                 $code.="        mov     $SZ*$i($inp),$T1\n";
222                 &ROUND_00_15($i,@ROT);
223                 unshift(@ROT,pop(@ROT));
224         }
225 $code.=<<___;
226         jmp     .Lrounds_16_xx
227 .align  16
228 .Ldata_order:
229 ___
230 } # 256
231         for($i=0;$i<16;$i++) {
232                 $code.="        mov     $SZ*$i($inp),$T1\n";
233                 $code.="        bswap   $T1\n";
234                 &ROUND_00_15($i,@ROT);
235                 unshift(@ROT,pop(@ROT));
236         }
237 $code.=<<___;
238         jmp     .Lrounds_16_xx
239 .align  16
240 .Lrounds_16_xx:
241 ___
242         for(;$i<32;$i++) {
243                 &ROUND_16_XX($i,@ROT);
244                 unshift(@ROT,pop(@ROT));
245         }
246
247 $code.=<<___;
248         cmp     \$$rounds,$round
249         jb      .Lrounds_16_xx
250
251         mov     $_ctx,$ctx
252         lea     16*$SZ($inp),$inp
253
254         add     $SZ*0($ctx),$A
255         add     $SZ*1($ctx),$B
256         add     $SZ*2($ctx),$C
257         add     $SZ*3($ctx),$D
258         add     $SZ*4($ctx),$E
259         add     $SZ*5($ctx),$F
260         add     $SZ*6($ctx),$G
261         add     $SZ*7($ctx),$H
262
263         cmp     $_end,$inp
264
265         mov     $A,$SZ*0($ctx)
266         mov     $B,$SZ*1($ctx)
267         mov     $C,$SZ*2($ctx)
268         mov     $D,$SZ*3($ctx)
269         mov     $E,$SZ*4($ctx)
270         mov     $F,$SZ*5($ctx)
271         mov     $G,$SZ*6($ctx)
272         mov     $H,$SZ*7($ctx)
273         jb      .Lloop
274
275         mov     $_rsp,%rsp
276         pop     %r15
277         pop     %r14
278         pop     %r13
279         pop     %r12
280         pop     %rbp
281         pop     %rbx
282
283         ret
284 .size   $func,.-$func
285 ___
286
287 if ($SZ==4) {
288 $code.=<<___;
289 .align  64
290 .type   $TABLE,\@object
291 $TABLE:
292         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
293         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
294         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
295         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
296         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
297         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
298         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
299         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
300         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
301         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
302         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
303         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
304         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
305         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
306         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
307         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
308 ___
309 } else {
310 $code.=<<___;
311 .align  64
312 .type   $TABLE,\@object
313 $TABLE:
314         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
315         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
316         .quad   0x3956c25bf348b538,0x59f111f1b605d019
317         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
318         .quad   0xd807aa98a3030242,0x12835b0145706fbe
319         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
320         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
321         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
322         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
323         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
324         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
325         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
326         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
327         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
328         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
329         .quad   0x06ca6351e003826f,0x142929670a0e6e70
330         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
331         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
332         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
333         .quad   0x81c2c92e47edaee6,0x92722c851482353b
334         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
335         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
336         .quad   0xd192e819d6ef5218,0xd69906245565a910
337         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
338         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
339         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
340         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
341         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
342         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
343         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
344         .quad   0x90befffa23631e28,0xa4506cebde82bde9
345         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
346         .quad   0xca273eceea26619c,0xd186b8c721c0c207
347         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
348         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
349         .quad   0x113f9804bef90dae,0x1b710b35131c471b
350         .quad   0x28db77f523047d84,0x32caab7b40c72493
351         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
352         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
353         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
354 ___
355 }
356
357 $code =~ s/\`([^\`]*)\`/eval $1/gem;
358 print $code;
359 close STDOUT;