ef517df64bf328bb4fabe9922e6e54e7af1762ae
[openssl.git] / crypto / sha / asm / sha512-armv8.pl
1 #! /usr/bin/env perl
2 # Copyright 2014-2016 The OpenSSL Project Authors. All Rights Reserved.
3 #
4 # Licensed under the OpenSSL license (the "License").  You may not use
5 # this file except in compliance with the License.  You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
8
9 #
10 # ====================================================================
11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12 # project. The module is, however, dual licensed under OpenSSL and
13 # CRYPTOGAMS licenses depending on where you obtain it. For further
14 # details see http://www.openssl.org/~appro/cryptogams/.
15 # ====================================================================
16 #
17 # SHA256/512 for ARMv8.
18 #
19 # Performance in cycles per processed byte and improvement coefficient
20 # over code generated with "default" compiler:
21 #
22 #               SHA256-hw       SHA256(*)       SHA512
23 # Apple A7      1.97            10.5 (+33%)     6.73 (-1%(**))
24 # Cortex-A53    2.38            15.5 (+115%)    10.0 (+150%(***))
25 # Cortex-A57    2.31            11.6 (+86%)     7.51 (+260%(***))
26 # Denver        2.01            10.5 (+26%)     6.70 (+8%)
27 # X-Gene                        20.0 (+100%)    12.8 (+300%(***))
28
29 # (*)   Software SHA256 results are of lesser relevance, presented
30 #       mostly for informational purposes.
31 # (**)  The result is a trade-off: it's possible to improve it by
32 #       10% (or by 1 cycle per round), but at the cost of 20% loss
33 #       on Cortex-A53 (or by 4 cycles per round).
34 # (***) Super-impressive coefficients over gcc-generated code are
35 #       indication of some compiler "pathology", most notably code
36 #       generated with -mgeneral-regs-only is significanty faster
37 #       and the gap is only 40-90%.
38
39 $flavour=shift;
40 $output=shift;
41
42 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
43 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
44 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
45 die "can't locate arm-xlate.pl";
46
47 open OUT,"| \"$^X\" $xlate $flavour $output";
48 *STDOUT=*OUT;
49
50 if ($output =~ /512/) {
51         $BITS=512;
52         $SZ=8;
53         @Sigma0=(28,34,39);
54         @Sigma1=(14,18,41);
55         @sigma0=(1,  8, 7);
56         @sigma1=(19,61, 6);
57         $rounds=80;
58         $reg_t="x";
59 } else {
60         $BITS=256;
61         $SZ=4;
62         @Sigma0=( 2,13,22);
63         @Sigma1=( 6,11,25);
64         @sigma0=( 7,18, 3);
65         @sigma1=(17,19,10);
66         $rounds=64;
67         $reg_t="w";
68 }
69
70 $func="sha${BITS}_block_data_order";
71
72 ($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
73
74 @X=map("$reg_t$_",(3..15,0..2));
75 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
76 ($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
77
78 sub BODY_00_xx {
79 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
80 my $j=($i+1)&15;
81 my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
82    $T0=@X[$i+3] if ($i<11);
83
84 $code.=<<___    if ($i<16);
85 #ifndef __ARMEB__
86         rev     @X[$i],@X[$i]                   // $i
87 #endif
88 ___
89 $code.=<<___    if ($i<13 && ($i&1));
90         ldp     @X[$i+1],@X[$i+2],[$inp],#2*$SZ
91 ___
92 $code.=<<___    if ($i==13);
93         ldp     @X[14],@X[15],[$inp]
94 ___
95 $code.=<<___    if ($i>=14);
96         ldr     @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
97 ___
98 $code.=<<___    if ($i>0 && $i<16);
99         add     $a,$a,$t1                       // h+=Sigma0(a)
100 ___
101 $code.=<<___    if ($i>=11);
102         str     @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
103 ___
104 # While ARMv8 specifies merged rotate-n-logical operation such as
105 # 'eor x,y,z,ror#n', it was found to negatively affect performance
106 # on Apple A7. The reason seems to be that it requires even 'y' to
107 # be available earlier. This means that such merged instruction is
108 # not necessarily best choice on critical path... On the other hand
109 # Cortex-A5x handles merged instructions much better than disjoint
110 # rotate and logical... See (**) footnote above.
111 $code.=<<___    if ($i<15);
112         ror     $t0,$e,#$Sigma1[0]
113         add     $h,$h,$t2                       // h+=K[i]
114         eor     $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
115         and     $t1,$f,$e
116         bic     $t2,$g,$e
117         add     $h,$h,@X[$i&15]                 // h+=X[i]
118         orr     $t1,$t1,$t2                     // Ch(e,f,g)
119         eor     $t2,$a,$b                       // a^b, b^c in next round
120         eor     $t0,$t0,$T0,ror#$Sigma1[1]      // Sigma1(e)
121         ror     $T0,$a,#$Sigma0[0]
122         add     $h,$h,$t1                       // h+=Ch(e,f,g)
123         eor     $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
124         add     $h,$h,$t0                       // h+=Sigma1(e)
125         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
126         add     $d,$d,$h                        // d+=h
127         eor     $t3,$t3,$b                      // Maj(a,b,c)
128         eor     $t1,$T0,$t1,ror#$Sigma0[1]      // Sigma0(a)
129         add     $h,$h,$t3                       // h+=Maj(a,b,c)
130         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
131         //add   $h,$h,$t1                       // h+=Sigma0(a)
132 ___
133 $code.=<<___    if ($i>=15);
134         ror     $t0,$e,#$Sigma1[0]
135         add     $h,$h,$t2                       // h+=K[i]
136         ror     $T1,@X[($j+1)&15],#$sigma0[0]
137         and     $t1,$f,$e
138         ror     $T2,@X[($j+14)&15],#$sigma1[0]
139         bic     $t2,$g,$e
140         ror     $T0,$a,#$Sigma0[0]
141         add     $h,$h,@X[$i&15]                 // h+=X[i]
142         eor     $t0,$t0,$e,ror#$Sigma1[1]
143         eor     $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
144         orr     $t1,$t1,$t2                     // Ch(e,f,g)
145         eor     $t2,$a,$b                       // a^b, b^c in next round
146         eor     $t0,$t0,$e,ror#$Sigma1[2]       // Sigma1(e)
147         eor     $T0,$T0,$a,ror#$Sigma0[1]
148         add     $h,$h,$t1                       // h+=Ch(e,f,g)
149         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
150         eor     $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
151         eor     $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]    // sigma0(X[i+1])
152         add     $h,$h,$t0                       // h+=Sigma1(e)
153         eor     $t3,$t3,$b                      // Maj(a,b,c)
154         eor     $t1,$T0,$a,ror#$Sigma0[2]       // Sigma0(a)
155         eor     $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]   // sigma1(X[i+14])
156         add     @X[$j],@X[$j],@X[($j+9)&15]
157         add     $d,$d,$h                        // d+=h
158         add     $h,$h,$t3                       // h+=Maj(a,b,c)
159         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
160         add     @X[$j],@X[$j],$T1
161         add     $h,$h,$t1                       // h+=Sigma0(a)
162         add     @X[$j],@X[$j],$T2
163 ___
164         ($t2,$t3)=($t3,$t2);
165 }
166
167 $code.=<<___;
168 #include "arm_arch.h"
169
170 .text
171
172 .extern OPENSSL_armcap_P
173 .globl  $func
174 .type   $func,%function
175 .align  6
176 $func:
177 ___
178 $code.=<<___    if ($SZ==4);
179 #ifdef  __ILP32__
180         ldrsw   x16,.LOPENSSL_armcap_P
181 #else
182         ldr     x16,.LOPENSSL_armcap_P
183 #endif
184         adr     x17,.LOPENSSL_armcap_P
185         add     x16,x16,x17
186         ldr     w16,[x16]
187         tst     w16,#ARMV8_SHA256
188         b.ne    .Lv8_entry
189 ___
190 $code.=<<___;
191         stp     x29,x30,[sp,#-128]!
192         add     x29,sp,#0
193
194         stp     x19,x20,[sp,#16]
195         stp     x21,x22,[sp,#32]
196         stp     x23,x24,[sp,#48]
197         stp     x25,x26,[sp,#64]
198         stp     x27,x28,[sp,#80]
199         sub     sp,sp,#4*$SZ
200
201         ldp     $A,$B,[$ctx]                            // load context
202         ldp     $C,$D,[$ctx,#2*$SZ]
203         ldp     $E,$F,[$ctx,#4*$SZ]
204         add     $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input
205         ldp     $G,$H,[$ctx,#6*$SZ]
206         adr     $Ktbl,.LK$BITS
207         stp     $ctx,$num,[x29,#96]
208
209 .Loop:
210         ldp     @X[0],@X[1],[$inp],#2*$SZ
211         ldr     $t2,[$Ktbl],#$SZ                        // *K++
212         eor     $t3,$B,$C                               // magic seed
213         str     $inp,[x29,#112]
214 ___
215 for ($i=0;$i<16;$i++)   { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
216 $code.=".Loop_16_xx:\n";
217 for (;$i<32;$i++)       { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
218 $code.=<<___;
219         cbnz    $t2,.Loop_16_xx
220
221         ldp     $ctx,$num,[x29,#96]
222         ldr     $inp,[x29,#112]
223         sub     $Ktbl,$Ktbl,#`$SZ*($rounds+1)`          // rewind
224
225         ldp     @X[0],@X[1],[$ctx]
226         ldp     @X[2],@X[3],[$ctx,#2*$SZ]
227         add     $inp,$inp,#14*$SZ                       // advance input pointer
228         ldp     @X[4],@X[5],[$ctx,#4*$SZ]
229         add     $A,$A,@X[0]
230         ldp     @X[6],@X[7],[$ctx,#6*$SZ]
231         add     $B,$B,@X[1]
232         add     $C,$C,@X[2]
233         add     $D,$D,@X[3]
234         stp     $A,$B,[$ctx]
235         add     $E,$E,@X[4]
236         add     $F,$F,@X[5]
237         stp     $C,$D,[$ctx,#2*$SZ]
238         add     $G,$G,@X[6]
239         add     $H,$H,@X[7]
240         cmp     $inp,$num
241         stp     $E,$F,[$ctx,#4*$SZ]
242         stp     $G,$H,[$ctx,#6*$SZ]
243         b.ne    .Loop
244
245         ldp     x19,x20,[x29,#16]
246         add     sp,sp,#4*$SZ
247         ldp     x21,x22,[x29,#32]
248         ldp     x23,x24,[x29,#48]
249         ldp     x25,x26,[x29,#64]
250         ldp     x27,x28,[x29,#80]
251         ldp     x29,x30,[sp],#128
252         ret
253 .size   $func,.-$func
254
255 .align  6
256 .type   .LK$BITS,%object
257 .LK$BITS:
258 ___
259 $code.=<<___ if ($SZ==8);
260         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
261         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
262         .quad   0x3956c25bf348b538,0x59f111f1b605d019
263         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
264         .quad   0xd807aa98a3030242,0x12835b0145706fbe
265         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
266         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
267         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
268         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
269         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
270         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
271         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
272         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
273         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
274         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
275         .quad   0x06ca6351e003826f,0x142929670a0e6e70
276         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
277         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
278         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
279         .quad   0x81c2c92e47edaee6,0x92722c851482353b
280         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
281         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
282         .quad   0xd192e819d6ef5218,0xd69906245565a910
283         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
284         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
285         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
286         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
287         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
288         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
289         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
290         .quad   0x90befffa23631e28,0xa4506cebde82bde9
291         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
292         .quad   0xca273eceea26619c,0xd186b8c721c0c207
293         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
294         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
295         .quad   0x113f9804bef90dae,0x1b710b35131c471b
296         .quad   0x28db77f523047d84,0x32caab7b40c72493
297         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
298         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
299         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
300         .quad   0       // terminator
301 ___
302 $code.=<<___ if ($SZ==4);
303         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
304         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
305         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
306         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
307         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
308         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
309         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
310         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
311         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
312         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
313         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
314         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
315         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
316         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
317         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
318         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
319         .long   0       //terminator
320 ___
321 $code.=<<___;
322 .size   .LK$BITS,.-.LK$BITS
323 .align  3
324 .LOPENSSL_armcap_P:
325 #ifdef  __ILP32__
326         .long   OPENSSL_armcap_P-.
327 #else
328         .quad   OPENSSL_armcap_P-.
329 #endif
330 .asciz  "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
331 .align  2
332 ___
333
334 if ($SZ==4) {
335 my $Ktbl="x3";
336
337 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
338 my @MSG=map("v$_.16b",(4..7));
339 my ($W0,$W1)=("v16.4s","v17.4s");
340 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
341
342 $code.=<<___;
343 .type   sha256_block_armv8,%function
344 .align  6
345 sha256_block_armv8:
346 .Lv8_entry:
347         stp             x29,x30,[sp,#-16]!
348         add             x29,sp,#0
349
350         ld1.32          {$ABCD,$EFGH},[$ctx]
351         adr             $Ktbl,.LK256
352
353 .Loop_hw:
354         ld1             {@MSG[0]-@MSG[3]},[$inp],#64
355         sub             $num,$num,#1
356         ld1.32          {$W0},[$Ktbl],#16
357         rev32           @MSG[0],@MSG[0]
358         rev32           @MSG[1],@MSG[1]
359         rev32           @MSG[2],@MSG[2]
360         rev32           @MSG[3],@MSG[3]
361         orr             $ABCD_SAVE,$ABCD,$ABCD          // offload
362         orr             $EFGH_SAVE,$EFGH,$EFGH
363 ___
364 for($i=0;$i<12;$i++) {
365 $code.=<<___;
366         ld1.32          {$W1},[$Ktbl],#16
367         add.i32         $W0,$W0,@MSG[0]
368         sha256su0       @MSG[0],@MSG[1]
369         orr             $abcd,$ABCD,$ABCD
370         sha256h         $ABCD,$EFGH,$W0
371         sha256h2        $EFGH,$abcd,$W0
372         sha256su1       @MSG[0],@MSG[2],@MSG[3]
373 ___
374         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
375 }
376 $code.=<<___;
377         ld1.32          {$W1},[$Ktbl],#16
378         add.i32         $W0,$W0,@MSG[0]
379         orr             $abcd,$ABCD,$ABCD
380         sha256h         $ABCD,$EFGH,$W0
381         sha256h2        $EFGH,$abcd,$W0
382
383         ld1.32          {$W0},[$Ktbl],#16
384         add.i32         $W1,$W1,@MSG[1]
385         orr             $abcd,$ABCD,$ABCD
386         sha256h         $ABCD,$EFGH,$W1
387         sha256h2        $EFGH,$abcd,$W1
388
389         ld1.32          {$W1},[$Ktbl]
390         add.i32         $W0,$W0,@MSG[2]
391         sub             $Ktbl,$Ktbl,#$rounds*$SZ-16     // rewind
392         orr             $abcd,$ABCD,$ABCD
393         sha256h         $ABCD,$EFGH,$W0
394         sha256h2        $EFGH,$abcd,$W0
395
396         add.i32         $W1,$W1,@MSG[3]
397         orr             $abcd,$ABCD,$ABCD
398         sha256h         $ABCD,$EFGH,$W1
399         sha256h2        $EFGH,$abcd,$W1
400
401         add.i32         $ABCD,$ABCD,$ABCD_SAVE
402         add.i32         $EFGH,$EFGH,$EFGH_SAVE
403
404         cbnz            $num,.Loop_hw
405
406         st1.32          {$ABCD,$EFGH},[$ctx]
407
408         ldr             x29,[sp],#16
409         ret
410 .size   sha256_block_armv8,.-sha256_block_armv8
411 ___
412 }
413
414 $code.=<<___;
415 .comm   OPENSSL_armcap_P,4,4
416 ___
417
418 {   my  %opcode = (
419         "sha256h"       => 0x5e004000,  "sha256h2"      => 0x5e005000,
420         "sha256su0"     => 0x5e282800,  "sha256su1"     => 0x5e006000   );
421
422     sub unsha256 {
423         my ($mnemonic,$arg)=@_;
424
425         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
426         &&
427         sprintf ".inst\t0x%08x\t//%s %s",
428                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
429                         $mnemonic,$arg;
430     }
431 }
432
433 foreach(split("\n",$code)) {
434
435         s/\`([^\`]*)\`/eval($1)/geo;
436
437         s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
438
439         s/\.\w?32\b//o          and s/\.16b/\.4s/go;
440         m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go;
441
442         print $_,"\n";
443 }
444
445 close STDOUT;