ac84ebb52e4f72c2acac664f37cf8635b8e61a6e
[openssl.git] / crypto / sha / asm / sha512-armv8.pl
1 #! /usr/bin/env perl
2 # Copyright 2014-2018 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 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
11 # project. The module is, however, dual licensed under OpenSSL and
12 # CRYPTOGAMS licenses depending on where you obtain it. For further
13 # details see http://www.openssl.org/~appro/cryptogams/.
14 #
15 # Permission to use under GPLv2 terms is granted.
16 # ====================================================================
17 #
18 # SHA256/512 for ARMv8.
19 #
20 # Performance in cycles per processed byte and improvement coefficient
21 # over code generated with "default" compiler:
22 #
23 #               SHA256-hw       SHA256(*)       SHA512
24 # Apple A7      1.97            10.5 (+33%)     6.73 (-1%(**))
25 # Cortex-A53    2.38            15.5 (+115%)    10.0 (+150%(***))
26 # Cortex-A57    2.31            11.6 (+86%)     7.51 (+260%(***))
27 # Denver        2.01            10.5 (+26%)     6.70 (+8%)
28 # X-Gene                        20.0 (+100%)    12.8 (+300%(***))
29 # Mongoose      2.36            13.0 (+50%)     8.36 (+33%)
30 # Kryo          1.92            17.4 (+30%)     11.2 (+8%)
31 #
32 # (*)   Software SHA256 results are of lesser relevance, presented
33 #       mostly for informational purposes.
34 # (**)  The result is a trade-off: it's possible to improve it by
35 #       10% (or by 1 cycle per round), but at the cost of 20% loss
36 #       on Cortex-A53 (or by 4 cycles per round).
37 # (***) Super-impressive coefficients over gcc-generated code are
38 #       indication of some compiler "pathology", most notably code
39 #       generated with -mgeneral-regs-only is significantly faster
40 #       and the gap is only 40-90%.
41 #
42 # October 2016.
43 #
44 # Originally it was reckoned that it makes no sense to implement NEON
45 # version of SHA256 for 64-bit processors. This is because performance
46 # improvement on most wide-spread Cortex-A5x processors was observed
47 # to be marginal, same on Cortex-A53 and ~10% on A57. But then it was
48 # observed that 32-bit NEON SHA256 performs significantly better than
49 # 64-bit scalar version on *some* of the more recent processors. As
50 # result 64-bit NEON version of SHA256 was added to provide best
51 # all-round performance. For example it executes ~30% faster on X-Gene
52 # and Mongoose. [For reference, NEON version of SHA512 is bound to
53 # deliver much less improvement, likely *negative* on Cortex-A5x.
54 # Which is why NEON support is limited to SHA256.]
55
56 $output=pop;
57 $flavour=pop;
58
59 if ($flavour && $flavour ne "void") {
60     $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
61     ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
62     ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
63     die "can't locate arm-xlate.pl";
64
65     open OUT,"| \"$^X\" $xlate $flavour $output";
66     *STDOUT=*OUT;
67 } else {
68     open STDOUT,">$output";
69 }
70
71 if ($output =~ /512/) {
72         $BITS=512;
73         $SZ=8;
74         @Sigma0=(28,34,39);
75         @Sigma1=(14,18,41);
76         @sigma0=(1,  8, 7);
77         @sigma1=(19,61, 6);
78         $rounds=80;
79         $reg_t="x";
80 } else {
81         $BITS=256;
82         $SZ=4;
83         @Sigma0=( 2,13,22);
84         @Sigma1=( 6,11,25);
85         @sigma0=( 7,18, 3);
86         @sigma1=(17,19,10);
87         $rounds=64;
88         $reg_t="w";
89 }
90
91 $func="sha${BITS}_block_data_order";
92
93 ($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
94
95 @X=map("$reg_t$_",(3..15,0..2));
96 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
97 ($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
98
99 sub BODY_00_xx {
100 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
101 my $j=($i+1)&15;
102 my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
103    $T0=@X[$i+3] if ($i<11);
104
105 $code.=<<___    if ($i<16);
106 #ifndef __AARCH64EB__
107         rev     @X[$i],@X[$i]                   // $i
108 #endif
109 ___
110 $code.=<<___    if ($i<13 && ($i&1));
111         ldp     @X[$i+1],@X[$i+2],[$inp],#2*$SZ
112 ___
113 $code.=<<___    if ($i==13);
114         ldp     @X[14],@X[15],[$inp]
115 ___
116 $code.=<<___    if ($i>=14);
117         ldr     @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
118 ___
119 $code.=<<___    if ($i>0 && $i<16);
120         add     $a,$a,$t1                       // h+=Sigma0(a)
121 ___
122 $code.=<<___    if ($i>=11);
123         str     @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
124 ___
125 # While ARMv8 specifies merged rotate-n-logical operation such as
126 # 'eor x,y,z,ror#n', it was found to negatively affect performance
127 # on Apple A7. The reason seems to be that it requires even 'y' to
128 # be available earlier. This means that such merged instruction is
129 # not necessarily best choice on critical path... On the other hand
130 # Cortex-A5x handles merged instructions much better than disjoint
131 # rotate and logical... See (**) footnote above.
132 $code.=<<___    if ($i<15);
133         ror     $t0,$e,#$Sigma1[0]
134         add     $h,$h,$t2                       // h+=K[i]
135         eor     $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
136         and     $t1,$f,$e
137         bic     $t2,$g,$e
138         add     $h,$h,@X[$i&15]                 // h+=X[i]
139         orr     $t1,$t1,$t2                     // Ch(e,f,g)
140         eor     $t2,$a,$b                       // a^b, b^c in next round
141         eor     $t0,$t0,$T0,ror#$Sigma1[1]      // Sigma1(e)
142         ror     $T0,$a,#$Sigma0[0]
143         add     $h,$h,$t1                       // h+=Ch(e,f,g)
144         eor     $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
145         add     $h,$h,$t0                       // h+=Sigma1(e)
146         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
147         add     $d,$d,$h                        // d+=h
148         eor     $t3,$t3,$b                      // Maj(a,b,c)
149         eor     $t1,$T0,$t1,ror#$Sigma0[1]      // Sigma0(a)
150         add     $h,$h,$t3                       // h+=Maj(a,b,c)
151         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
152         //add   $h,$h,$t1                       // h+=Sigma0(a)
153 ___
154 $code.=<<___    if ($i>=15);
155         ror     $t0,$e,#$Sigma1[0]
156         add     $h,$h,$t2                       // h+=K[i]
157         ror     $T1,@X[($j+1)&15],#$sigma0[0]
158         and     $t1,$f,$e
159         ror     $T2,@X[($j+14)&15],#$sigma1[0]
160         bic     $t2,$g,$e
161         ror     $T0,$a,#$Sigma0[0]
162         add     $h,$h,@X[$i&15]                 // h+=X[i]
163         eor     $t0,$t0,$e,ror#$Sigma1[1]
164         eor     $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
165         orr     $t1,$t1,$t2                     // Ch(e,f,g)
166         eor     $t2,$a,$b                       // a^b, b^c in next round
167         eor     $t0,$t0,$e,ror#$Sigma1[2]       // Sigma1(e)
168         eor     $T0,$T0,$a,ror#$Sigma0[1]
169         add     $h,$h,$t1                       // h+=Ch(e,f,g)
170         and     $t3,$t3,$t2                     // (b^c)&=(a^b)
171         eor     $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
172         eor     $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]    // sigma0(X[i+1])
173         add     $h,$h,$t0                       // h+=Sigma1(e)
174         eor     $t3,$t3,$b                      // Maj(a,b,c)
175         eor     $t1,$T0,$a,ror#$Sigma0[2]       // Sigma0(a)
176         eor     $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]   // sigma1(X[i+14])
177         add     @X[$j],@X[$j],@X[($j+9)&15]
178         add     $d,$d,$h                        // d+=h
179         add     $h,$h,$t3                       // h+=Maj(a,b,c)
180         ldr     $t3,[$Ktbl],#$SZ                // *K++, $t2 in next round
181         add     @X[$j],@X[$j],$T1
182         add     $h,$h,$t1                       // h+=Sigma0(a)
183         add     @X[$j],@X[$j],$T2
184 ___
185         ($t2,$t3)=($t3,$t2);
186 }
187
188 $code.=<<___;
189 #ifndef __KERNEL__
190 # include "arm_arch.h"
191 #endif
192
193 .text
194
195 .extern OPENSSL_armcap_P
196 .globl  $func
197 .type   $func,%function
198 .align  6
199 $func:
200 #ifndef __KERNEL__
201 # ifdef __ILP32__
202         ldrsw   x16,.LOPENSSL_armcap_P
203 # else
204         ldr     x16,.LOPENSSL_armcap_P
205 # endif
206         adr     x17,.LOPENSSL_armcap_P
207         add     x16,x16,x17
208         ldr     w16,[x16]
209 ___
210 $code.=<<___    if ($SZ==4);
211         tst     w16,#ARMV8_SHA256
212         b.ne    .Lv8_entry
213         tst     w16,#ARMV7_NEON
214         b.ne    .Lneon_entry
215 ___
216 $code.=<<___    if ($SZ==8);
217         tst     w16,#ARMV8_SHA512
218         b.ne    .Lv8_entry
219 ___
220 $code.=<<___;
221 #endif
222         stp     x29,x30,[sp,#-128]!
223         add     x29,sp,#0
224
225         stp     x19,x20,[sp,#16]
226         stp     x21,x22,[sp,#32]
227         stp     x23,x24,[sp,#48]
228         stp     x25,x26,[sp,#64]
229         stp     x27,x28,[sp,#80]
230         sub     sp,sp,#4*$SZ
231
232         ldp     $A,$B,[$ctx]                            // load context
233         ldp     $C,$D,[$ctx,#2*$SZ]
234         ldp     $E,$F,[$ctx,#4*$SZ]
235         add     $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input
236         ldp     $G,$H,[$ctx,#6*$SZ]
237         adr     $Ktbl,.LK$BITS
238         stp     $ctx,$num,[x29,#96]
239
240 .Loop:
241         ldp     @X[0],@X[1],[$inp],#2*$SZ
242         ldr     $t2,[$Ktbl],#$SZ                        // *K++
243         eor     $t3,$B,$C                               // magic seed
244         str     $inp,[x29,#112]
245 ___
246 for ($i=0;$i<16;$i++)   { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
247 $code.=".Loop_16_xx:\n";
248 for (;$i<32;$i++)       { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
249 $code.=<<___;
250         cbnz    $t2,.Loop_16_xx
251
252         ldp     $ctx,$num,[x29,#96]
253         ldr     $inp,[x29,#112]
254         sub     $Ktbl,$Ktbl,#`$SZ*($rounds+1)`          // rewind
255
256         ldp     @X[0],@X[1],[$ctx]
257         ldp     @X[2],@X[3],[$ctx,#2*$SZ]
258         add     $inp,$inp,#14*$SZ                       // advance input pointer
259         ldp     @X[4],@X[5],[$ctx,#4*$SZ]
260         add     $A,$A,@X[0]
261         ldp     @X[6],@X[7],[$ctx,#6*$SZ]
262         add     $B,$B,@X[1]
263         add     $C,$C,@X[2]
264         add     $D,$D,@X[3]
265         stp     $A,$B,[$ctx]
266         add     $E,$E,@X[4]
267         add     $F,$F,@X[5]
268         stp     $C,$D,[$ctx,#2*$SZ]
269         add     $G,$G,@X[6]
270         add     $H,$H,@X[7]
271         cmp     $inp,$num
272         stp     $E,$F,[$ctx,#4*$SZ]
273         stp     $G,$H,[$ctx,#6*$SZ]
274         b.ne    .Loop
275
276         ldp     x19,x20,[x29,#16]
277         add     sp,sp,#4*$SZ
278         ldp     x21,x22,[x29,#32]
279         ldp     x23,x24,[x29,#48]
280         ldp     x25,x26,[x29,#64]
281         ldp     x27,x28,[x29,#80]
282         ldp     x29,x30,[sp],#128
283         ret
284 .size   $func,.-$func
285
286 .align  6
287 .type   .LK$BITS,%object
288 .LK$BITS:
289 ___
290 $code.=<<___ if ($SZ==8);
291         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
292         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
293         .quad   0x3956c25bf348b538,0x59f111f1b605d019
294         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
295         .quad   0xd807aa98a3030242,0x12835b0145706fbe
296         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
297         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
298         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
299         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
300         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
301         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
302         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
303         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
304         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
305         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
306         .quad   0x06ca6351e003826f,0x142929670a0e6e70
307         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
308         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
309         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
310         .quad   0x81c2c92e47edaee6,0x92722c851482353b
311         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
312         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
313         .quad   0xd192e819d6ef5218,0xd69906245565a910
314         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
315         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
316         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
317         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
318         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
319         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
320         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
321         .quad   0x90befffa23631e28,0xa4506cebde82bde9
322         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
323         .quad   0xca273eceea26619c,0xd186b8c721c0c207
324         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
325         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
326         .quad   0x113f9804bef90dae,0x1b710b35131c471b
327         .quad   0x28db77f523047d84,0x32caab7b40c72493
328         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
329         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
330         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
331         .quad   0       // terminator
332 ___
333 $code.=<<___ if ($SZ==4);
334         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
335         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
336         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
337         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
338         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
339         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
340         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
341         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
342         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
343         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
344         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
345         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
346         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
347         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
348         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
349         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
350         .long   0       //terminator
351 ___
352 $code.=<<___;
353 .size   .LK$BITS,.-.LK$BITS
354 #ifndef __KERNEL__
355 .align  3
356 .LOPENSSL_armcap_P:
357 # ifdef __ILP32__
358         .long   OPENSSL_armcap_P-.
359 # else
360         .quad   OPENSSL_armcap_P-.
361 # endif
362 #endif
363 .asciz  "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
364 .align  2
365 ___
366
367 if ($SZ==4) {
368 my $Ktbl="x3";
369
370 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
371 my @MSG=map("v$_.16b",(4..7));
372 my ($W0,$W1)=("v16.4s","v17.4s");
373 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
374
375 $code.=<<___;
376 #ifndef __KERNEL__
377 .type   sha256_block_armv8,%function
378 .align  6
379 sha256_block_armv8:
380 .Lv8_entry:
381         stp             x29,x30,[sp,#-16]!
382         add             x29,sp,#0
383
384         ld1.32          {$ABCD,$EFGH},[$ctx]
385         adr             $Ktbl,.LK256
386
387 .Loop_hw:
388         ld1             {@MSG[0]-@MSG[3]},[$inp],#64
389         sub             $num,$num,#1
390         ld1.32          {$W0},[$Ktbl],#16
391         rev32           @MSG[0],@MSG[0]
392         rev32           @MSG[1],@MSG[1]
393         rev32           @MSG[2],@MSG[2]
394         rev32           @MSG[3],@MSG[3]
395         orr             $ABCD_SAVE,$ABCD,$ABCD          // offload
396         orr             $EFGH_SAVE,$EFGH,$EFGH
397 ___
398 for($i=0;$i<12;$i++) {
399 $code.=<<___;
400         ld1.32          {$W1},[$Ktbl],#16
401         add.i32         $W0,$W0,@MSG[0]
402         sha256su0       @MSG[0],@MSG[1]
403         orr             $abcd,$ABCD,$ABCD
404         sha256h         $ABCD,$EFGH,$W0
405         sha256h2        $EFGH,$abcd,$W0
406         sha256su1       @MSG[0],@MSG[2],@MSG[3]
407 ___
408         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
409 }
410 $code.=<<___;
411         ld1.32          {$W1},[$Ktbl],#16
412         add.i32         $W0,$W0,@MSG[0]
413         orr             $abcd,$ABCD,$ABCD
414         sha256h         $ABCD,$EFGH,$W0
415         sha256h2        $EFGH,$abcd,$W0
416
417         ld1.32          {$W0},[$Ktbl],#16
418         add.i32         $W1,$W1,@MSG[1]
419         orr             $abcd,$ABCD,$ABCD
420         sha256h         $ABCD,$EFGH,$W1
421         sha256h2        $EFGH,$abcd,$W1
422
423         ld1.32          {$W1},[$Ktbl]
424         add.i32         $W0,$W0,@MSG[2]
425         sub             $Ktbl,$Ktbl,#$rounds*$SZ-16     // rewind
426         orr             $abcd,$ABCD,$ABCD
427         sha256h         $ABCD,$EFGH,$W0
428         sha256h2        $EFGH,$abcd,$W0
429
430         add.i32         $W1,$W1,@MSG[3]
431         orr             $abcd,$ABCD,$ABCD
432         sha256h         $ABCD,$EFGH,$W1
433         sha256h2        $EFGH,$abcd,$W1
434
435         add.i32         $ABCD,$ABCD,$ABCD_SAVE
436         add.i32         $EFGH,$EFGH,$EFGH_SAVE
437
438         cbnz            $num,.Loop_hw
439
440         st1.32          {$ABCD,$EFGH},[$ctx]
441
442         ldr             x29,[sp],#16
443         ret
444 .size   sha256_block_armv8,.-sha256_block_armv8
445 #endif
446 ___
447 }
448
449 if ($SZ==4) {   ######################################### NEON stuff #
450 # You'll surely note a lot of similarities with sha256-armv4 module,
451 # and of course it's not a coincidence. sha256-armv4 was used as
452 # initial template, but was adapted for ARMv8 instruction set and
453 # extensively re-tuned for all-round performance.
454
455 my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
456 my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
457 my $Ktbl="x16";
458 my $Xfer="x17";
459 my @X = map("q$_",(0..3));
460 my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
461 my $j=0;
462
463 sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
464 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
465   my $arg = pop;
466     $arg = "#$arg" if ($arg*1 eq $arg);
467     $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
468 }
469
470 sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
471 sub Dlo     { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
472 sub Dhi     { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
473
474 sub Xupdate()
475 { use integer;
476   my $body = shift;
477   my @insns = (&$body,&$body,&$body,&$body);
478   my ($a,$b,$c,$d,$e,$f,$g,$h);
479
480         &ext_8          ($T0,@X[0],@X[1],4);    # X[1..4]
481          eval(shift(@insns));
482          eval(shift(@insns));
483          eval(shift(@insns));
484         &ext_8          ($T3,@X[2],@X[3],4);    # X[9..12]
485          eval(shift(@insns));
486          eval(shift(@insns));
487         &mov            (&Dscalar($T7),&Dhi(@X[3]));    # X[14..15]
488          eval(shift(@insns));
489          eval(shift(@insns));
490         &ushr_32        ($T2,$T0,$sigma0[0]);
491          eval(shift(@insns));
492         &ushr_32        ($T1,$T0,$sigma0[2]);
493          eval(shift(@insns));
494         &add_32         (@X[0],@X[0],$T3);      # X[0..3] += X[9..12]
495          eval(shift(@insns));
496         &sli_32         ($T2,$T0,32-$sigma0[0]);
497          eval(shift(@insns));
498          eval(shift(@insns));
499         &ushr_32        ($T3,$T0,$sigma0[1]);
500          eval(shift(@insns));
501          eval(shift(@insns));
502         &eor_8          ($T1,$T1,$T2);
503          eval(shift(@insns));
504          eval(shift(@insns));
505         &sli_32         ($T3,$T0,32-$sigma0[1]);
506          eval(shift(@insns));
507          eval(shift(@insns));
508           &ushr_32      ($T4,$T7,$sigma1[0]);
509          eval(shift(@insns));
510          eval(shift(@insns));
511         &eor_8          ($T1,$T1,$T3);          # sigma0(X[1..4])
512          eval(shift(@insns));
513          eval(shift(@insns));
514           &sli_32       ($T4,$T7,32-$sigma1[0]);
515          eval(shift(@insns));
516          eval(shift(@insns));
517           &ushr_32      ($T5,$T7,$sigma1[2]);
518          eval(shift(@insns));
519          eval(shift(@insns));
520           &ushr_32      ($T3,$T7,$sigma1[1]);
521          eval(shift(@insns));
522          eval(shift(@insns));
523         &add_32         (@X[0],@X[0],$T1);      # X[0..3] += sigma0(X[1..4])
524          eval(shift(@insns));
525          eval(shift(@insns));
526           &sli_u32      ($T3,$T7,32-$sigma1[1]);
527          eval(shift(@insns));
528          eval(shift(@insns));
529           &eor_8        ($T5,$T5,$T4);
530          eval(shift(@insns));
531          eval(shift(@insns));
532          eval(shift(@insns));
533           &eor_8        ($T5,$T5,$T3);          # sigma1(X[14..15])
534          eval(shift(@insns));
535          eval(shift(@insns));
536          eval(shift(@insns));
537         &add_32         (@X[0],@X[0],$T5);      # X[0..1] += sigma1(X[14..15])
538          eval(shift(@insns));
539          eval(shift(@insns));
540          eval(shift(@insns));
541           &ushr_32      ($T6,@X[0],$sigma1[0]);
542          eval(shift(@insns));
543           &ushr_32      ($T7,@X[0],$sigma1[2]);
544          eval(shift(@insns));
545          eval(shift(@insns));
546           &sli_32       ($T6,@X[0],32-$sigma1[0]);
547          eval(shift(@insns));
548           &ushr_32      ($T5,@X[0],$sigma1[1]);
549          eval(shift(@insns));
550          eval(shift(@insns));
551           &eor_8        ($T7,$T7,$T6);
552          eval(shift(@insns));
553          eval(shift(@insns));
554           &sli_32       ($T5,@X[0],32-$sigma1[1]);
555          eval(shift(@insns));
556          eval(shift(@insns));
557         &ld1_32         ("{$T0}","[$Ktbl], #16");
558          eval(shift(@insns));
559           &eor_8        ($T7,$T7,$T5);          # sigma1(X[16..17])
560          eval(shift(@insns));
561          eval(shift(@insns));
562         &eor_8          ($T5,$T5,$T5);
563          eval(shift(@insns));
564          eval(shift(@insns));
565         &mov            (&Dhi($T5), &Dlo($T7));
566          eval(shift(@insns));
567          eval(shift(@insns));
568          eval(shift(@insns));
569         &add_32         (@X[0],@X[0],$T5);      # X[2..3] += sigma1(X[16..17])
570          eval(shift(@insns));
571          eval(shift(@insns));
572          eval(shift(@insns));
573         &add_32         ($T0,$T0,@X[0]);
574          while($#insns>=1) { eval(shift(@insns)); }
575         &st1_32         ("{$T0}","[$Xfer], #16");
576          eval(shift(@insns));
577
578         push(@X,shift(@X));             # "rotate" X[]
579 }
580
581 sub Xpreload()
582 { use integer;
583   my $body = shift;
584   my @insns = (&$body,&$body,&$body,&$body);
585   my ($a,$b,$c,$d,$e,$f,$g,$h);
586
587          eval(shift(@insns));
588          eval(shift(@insns));
589         &ld1_8          ("{@X[0]}","[$inp],#16");
590          eval(shift(@insns));
591          eval(shift(@insns));
592         &ld1_32         ("{$T0}","[$Ktbl],#16");
593          eval(shift(@insns));
594          eval(shift(@insns));
595          eval(shift(@insns));
596          eval(shift(@insns));
597         &rev32          (@X[0],@X[0]);
598          eval(shift(@insns));
599          eval(shift(@insns));
600          eval(shift(@insns));
601          eval(shift(@insns));
602         &add_32         ($T0,$T0,@X[0]);
603          foreach (@insns) { eval; }     # remaining instructions
604         &st1_32         ("{$T0}","[$Xfer], #16");
605
606         push(@X,shift(@X));             # "rotate" X[]
607 }
608
609 sub body_00_15 () {
610         (
611         '($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
612         '&add   ($h,$h,$t1)',                   # h+=X[i]+K[i]
613         '&add   ($a,$a,$t4);'.                  # h+=Sigma0(a) from the past
614         '&and   ($t1,$f,$e)',
615         '&bic   ($t4,$g,$e)',
616         '&eor   ($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
617         '&add   ($a,$a,$t2)',                   # h+=Maj(a,b,c) from the past
618         '&orr   ($t1,$t1,$t4)',                 # Ch(e,f,g)
619         '&eor   ($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))',   # Sigma1(e)
620         '&eor   ($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
621         '&add   ($h,$h,$t1)',                   # h+=Ch(e,f,g)
622         '&ror   ($t0,$t0,"#$Sigma1[0]")',
623         '&eor   ($t2,$a,$b)',                   # a^b, b^c in next round
624         '&eor   ($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))',   # Sigma0(a)
625         '&add   ($h,$h,$t0)',                   # h+=Sigma1(e)
626         '&ldr   ($t1,sprintf "[sp,#%d]",4*(($j+1)&15))  if (($j&15)!=15);'.
627         '&ldr   ($t1,"[$Ktbl]")                         if ($j==15);'.
628         '&and   ($t3,$t3,$t2)',                 # (b^c)&=(a^b)
629         '&ror   ($t4,$t4,"#$Sigma0[0]")',
630         '&add   ($d,$d,$h)',                    # d+=h
631         '&eor   ($t3,$t3,$b)',                  # Maj(a,b,c)
632         '$j++;  unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
633         )
634 }
635
636 $code.=<<___;
637 #ifdef  __KERNEL__
638 .globl  sha256_block_neon
639 #endif
640 .type   sha256_block_neon,%function
641 .align  4
642 sha256_block_neon:
643 .Lneon_entry:
644         stp     x29, x30, [sp, #-16]!
645         mov     x29, sp
646         sub     sp,sp,#16*4
647
648         adr     $Ktbl,.LK256
649         add     $num,$inp,$num,lsl#6    // len to point at the end of inp
650
651         ld1.8   {@X[0]},[$inp], #16
652         ld1.8   {@X[1]},[$inp], #16
653         ld1.8   {@X[2]},[$inp], #16
654         ld1.8   {@X[3]},[$inp], #16
655         ld1.32  {$T0},[$Ktbl], #16
656         ld1.32  {$T1},[$Ktbl], #16
657         ld1.32  {$T2},[$Ktbl], #16
658         ld1.32  {$T3},[$Ktbl], #16
659         rev32   @X[0],@X[0]             // yes, even on
660         rev32   @X[1],@X[1]             // big-endian
661         rev32   @X[2],@X[2]
662         rev32   @X[3],@X[3]
663         mov     $Xfer,sp
664         add.32  $T0,$T0,@X[0]
665         add.32  $T1,$T1,@X[1]
666         add.32  $T2,$T2,@X[2]
667         st1.32  {$T0-$T1},[$Xfer], #32
668         add.32  $T3,$T3,@X[3]
669         st1.32  {$T2-$T3},[$Xfer]
670         sub     $Xfer,$Xfer,#32
671
672         ldp     $A,$B,[$ctx]
673         ldp     $C,$D,[$ctx,#8]
674         ldp     $E,$F,[$ctx,#16]
675         ldp     $G,$H,[$ctx,#24]
676         ldr     $t1,[sp,#0]
677         mov     $t2,wzr
678         eor     $t3,$B,$C
679         mov     $t4,wzr
680         b       .L_00_48
681
682 .align  4
683 .L_00_48:
684 ___
685         &Xupdate(\&body_00_15);
686         &Xupdate(\&body_00_15);
687         &Xupdate(\&body_00_15);
688         &Xupdate(\&body_00_15);
689 $code.=<<___;
690         cmp     $t1,#0                          // check for K256 terminator
691         ldr     $t1,[sp,#0]
692         sub     $Xfer,$Xfer,#64
693         bne     .L_00_48
694
695         sub     $Ktbl,$Ktbl,#256                // rewind $Ktbl
696         cmp     $inp,$num
697         mov     $Xfer, #64
698         csel    $Xfer, $Xfer, xzr, eq
699         sub     $inp,$inp,$Xfer                 // avoid SEGV
700         mov     $Xfer,sp
701 ___
702         &Xpreload(\&body_00_15);
703         &Xpreload(\&body_00_15);
704         &Xpreload(\&body_00_15);
705         &Xpreload(\&body_00_15);
706 $code.=<<___;
707         add     $A,$A,$t4                       // h+=Sigma0(a) from the past
708         ldp     $t0,$t1,[$ctx,#0]
709         add     $A,$A,$t2                       // h+=Maj(a,b,c) from the past
710         ldp     $t2,$t3,[$ctx,#8]
711         add     $A,$A,$t0                       // accumulate
712         add     $B,$B,$t1
713         ldp     $t0,$t1,[$ctx,#16]
714         add     $C,$C,$t2
715         add     $D,$D,$t3
716         ldp     $t2,$t3,[$ctx,#24]
717         add     $E,$E,$t0
718         add     $F,$F,$t1
719          ldr    $t1,[sp,#0]
720         stp     $A,$B,[$ctx,#0]
721         add     $G,$G,$t2
722          mov    $t2,wzr
723         stp     $C,$D,[$ctx,#8]
724         add     $H,$H,$t3
725         stp     $E,$F,[$ctx,#16]
726          eor    $t3,$B,$C
727         stp     $G,$H,[$ctx,#24]
728          mov    $t4,wzr
729          mov    $Xfer,sp
730         b.ne    .L_00_48
731
732         ldr     x29,[x29]
733         add     sp,sp,#16*4+16
734         ret
735 .size   sha256_block_neon,.-sha256_block_neon
736 ___
737 }
738
739 if ($SZ==8) {
740 my $Ktbl="x3";
741
742 my @H = map("v$_.16b",(0..4));
743 my ($fg,$de,$m9_10)=map("v$_.16b",(5..7));
744 my @MSG=map("v$_.16b",(16..23));
745 my ($W0,$W1)=("v24.2d","v25.2d");
746 my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29));
747
748 $code.=<<___;
749 #ifndef __KERNEL__
750 .type   sha512_block_armv8,%function
751 .align  6
752 sha512_block_armv8:
753 .Lv8_entry:
754         stp             x29,x30,[sp,#-16]!
755         add             x29,sp,#0
756
757         ld1             {@MSG[0]-@MSG[3]},[$inp],#64    // load input
758         ld1             {@MSG[4]-@MSG[7]},[$inp],#64
759
760         ld1.64          {@H[0]-@H[3]},[$ctx]            // load context
761         adr             $Ktbl,.LK512
762
763         rev64           @MSG[0],@MSG[0]
764         rev64           @MSG[1],@MSG[1]
765         rev64           @MSG[2],@MSG[2]
766         rev64           @MSG[3],@MSG[3]
767         rev64           @MSG[4],@MSG[4]
768         rev64           @MSG[5],@MSG[5]
769         rev64           @MSG[6],@MSG[6]
770         rev64           @MSG[7],@MSG[7]
771         b               .Loop_hw
772
773 .align  4
774 .Loop_hw:
775         ld1.64          {$W0},[$Ktbl],#16
776         subs            $num,$num,#1
777         sub             x4,$inp,#128
778         orr             $AB,@H[0],@H[0]                 // offload
779         orr             $CD,@H[1],@H[1]
780         orr             $EF,@H[2],@H[2]
781         orr             $GH,@H[3],@H[3]
782         csel            $inp,$inp,x4,ne                 // conditional rewind
783 ___
784 for($i=0;$i<32;$i++) {
785 $code.=<<___;
786         add.i64         $W0,$W0,@MSG[0]
787         ld1.64          {$W1},[$Ktbl],#16
788         ext             $W0,$W0,$W0,#8
789         ext             $fg,@H[2],@H[3],#8
790         ext             $de,@H[1],@H[2],#8
791         add.i64         @H[3],@H[3],$W0                 // "T1 + H + K512[i]"
792          sha512su0      @MSG[0],@MSG[1]
793          ext            $m9_10,@MSG[4],@MSG[5],#8
794         sha512h         @H[3],$fg,$de
795          sha512su1      @MSG[0],@MSG[7],$m9_10
796         add.i64         @H[4],@H[1],@H[3]               // "D + T1"
797         sha512h2        @H[3],$H[1],@H[0]
798 ___
799         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
800         @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
801 }
802 for(;$i<40;$i++) {
803 $code.=<<___    if ($i<39);
804         ld1.64          {$W1},[$Ktbl],#16
805 ___
806 $code.=<<___    if ($i==39);
807         sub             $Ktbl,$Ktbl,#$rounds*$SZ        // rewind
808 ___
809 $code.=<<___;
810         add.i64         $W0,$W0,@MSG[0]
811          ld1            {@MSG[0]},[$inp],#16            // load next input
812         ext             $W0,$W0,$W0,#8
813         ext             $fg,@H[2],@H[3],#8
814         ext             $de,@H[1],@H[2],#8
815         add.i64         @H[3],@H[3],$W0                 // "T1 + H + K512[i]"
816         sha512h         @H[3],$fg,$de
817          rev64          @MSG[0],@MSG[0]
818         add.i64         @H[4],@H[1],@H[3]               // "D + T1"
819         sha512h2        @H[3],$H[1],@H[0]
820 ___
821         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
822         @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
823 }
824 $code.=<<___;
825         add.i64         @H[0],@H[0],$AB                 // accumulate
826         add.i64         @H[1],@H[1],$CD
827         add.i64         @H[2],@H[2],$EF
828         add.i64         @H[3],@H[3],$GH
829
830         cbnz            $num,.Loop_hw
831
832         st1.64          {@H[0]-@H[3]},[$ctx]            // store context
833
834         ldr             x29,[sp],#16
835         ret
836 .size   sha512_block_armv8,.-sha512_block_armv8
837 #endif
838 ___
839 }
840
841 $code.=<<___;
842 #ifndef __KERNEL__
843 .comm   OPENSSL_armcap_P,4,4
844 #endif
845 ___
846
847 {   my  %opcode = (
848         "sha256h"       => 0x5e004000,  "sha256h2"      => 0x5e005000,
849         "sha256su0"     => 0x5e282800,  "sha256su1"     => 0x5e006000   );
850
851     sub unsha256 {
852         my ($mnemonic,$arg)=@_;
853
854         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
855         &&
856         sprintf ".inst\t0x%08x\t//%s %s",
857                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
858                         $mnemonic,$arg;
859     }
860 }
861
862 {   my  %opcode = (
863         "sha512h"       => 0xce608000,  "sha512h2"      => 0xce608400,
864         "sha512su0"     => 0xcec08000,  "sha512su1"     => 0xce608800   );
865
866     sub unsha512 {
867         my ($mnemonic,$arg)=@_;
868
869         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
870         &&
871         sprintf ".inst\t0x%08x\t//%s %s",
872                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
873                         $mnemonic,$arg;
874     }
875 }
876
877 open SELF,$0;
878 while(<SELF>) {
879         next if (/^#!/);
880         last if (!s/^#/\/\// and !/^$/);
881         print;
882 }
883 close SELF;
884
885 foreach(split("\n",$code)) {
886
887         s/\`([^\`]*)\`/eval($1)/ge;
888
889         s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge   or
890         s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
891
892         s/\bq([0-9]+)\b/v$1.16b/g;              # old->new registers
893
894         s/\.[ui]?8(\s)/$1/;
895         s/\.\w?64\b//           and s/\.16b/\.2d/g      or
896         s/\.\w?32\b//           and s/\.16b/\.4s/g;
897         m/\bext\b/              and s/\.2d/\.16b/g      or
898         m/(ld|st)1[^\[]+\[0\]/  and s/\.4s/\.s/g;
899
900         print $_,"\n";
901 }
902
903 close STDOUT;