60df899f328d9c9f93a1497cf268f6312ced1221
[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 $code.=<<___;
221 #endif
222         .inst   0xd503233f                              // paciasp
223         stp     x29,x30,[sp,#-128]!
224         add     x29,sp,#0
225
226         stp     x19,x20,[sp,#16]
227         stp     x21,x22,[sp,#32]
228         stp     x23,x24,[sp,#48]
229         stp     x25,x26,[sp,#64]
230         stp     x27,x28,[sp,#80]
231         sub     sp,sp,#4*$SZ
232
233         ldp     $A,$B,[$ctx]                            // load context
234         ldp     $C,$D,[$ctx,#2*$SZ]
235         ldp     $E,$F,[$ctx,#4*$SZ]
236         add     $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input
237         ldp     $G,$H,[$ctx,#6*$SZ]
238         adr     $Ktbl,.LK$BITS
239         stp     $ctx,$num,[x29,#96]
240
241 .Loop:
242         ldp     @X[0],@X[1],[$inp],#2*$SZ
243         ldr     $t2,[$Ktbl],#$SZ                        // *K++
244         eor     $t3,$B,$C                               // magic seed
245         str     $inp,[x29,#112]
246 ___
247 for ($i=0;$i<16;$i++)   { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
248 $code.=".Loop_16_xx:\n";
249 for (;$i<32;$i++)       { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
250 $code.=<<___;
251         cbnz    $t2,.Loop_16_xx
252
253         ldp     $ctx,$num,[x29,#96]
254         ldr     $inp,[x29,#112]
255         sub     $Ktbl,$Ktbl,#`$SZ*($rounds+1)`          // rewind
256
257         ldp     @X[0],@X[1],[$ctx]
258         ldp     @X[2],@X[3],[$ctx,#2*$SZ]
259         add     $inp,$inp,#14*$SZ                       // advance input pointer
260         ldp     @X[4],@X[5],[$ctx,#4*$SZ]
261         add     $A,$A,@X[0]
262         ldp     @X[6],@X[7],[$ctx,#6*$SZ]
263         add     $B,$B,@X[1]
264         add     $C,$C,@X[2]
265         add     $D,$D,@X[3]
266         stp     $A,$B,[$ctx]
267         add     $E,$E,@X[4]
268         add     $F,$F,@X[5]
269         stp     $C,$D,[$ctx,#2*$SZ]
270         add     $G,$G,@X[6]
271         add     $H,$H,@X[7]
272         cmp     $inp,$num
273         stp     $E,$F,[$ctx,#4*$SZ]
274         stp     $G,$H,[$ctx,#6*$SZ]
275         b.ne    .Loop
276
277         ldp     x19,x20,[x29,#16]
278         add     sp,sp,#4*$SZ
279         ldp     x21,x22,[x29,#32]
280         ldp     x23,x24,[x29,#48]
281         ldp     x25,x26,[x29,#64]
282         ldp     x27,x28,[x29,#80]
283         ldp     x29,x30,[sp],#128
284         .inst   0xd50323bf                              // autiasp
285         ret
286 .size   $func,.-$func
287
288 .align  6
289 .type   .LK$BITS,%object
290 .LK$BITS:
291 ___
292 $code.=<<___ if ($SZ==8);
293         .quad   0x428a2f98d728ae22,0x7137449123ef65cd
294         .quad   0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
295         .quad   0x3956c25bf348b538,0x59f111f1b605d019
296         .quad   0x923f82a4af194f9b,0xab1c5ed5da6d8118
297         .quad   0xd807aa98a3030242,0x12835b0145706fbe
298         .quad   0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
299         .quad   0x72be5d74f27b896f,0x80deb1fe3b1696b1
300         .quad   0x9bdc06a725c71235,0xc19bf174cf692694
301         .quad   0xe49b69c19ef14ad2,0xefbe4786384f25e3
302         .quad   0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
303         .quad   0x2de92c6f592b0275,0x4a7484aa6ea6e483
304         .quad   0x5cb0a9dcbd41fbd4,0x76f988da831153b5
305         .quad   0x983e5152ee66dfab,0xa831c66d2db43210
306         .quad   0xb00327c898fb213f,0xbf597fc7beef0ee4
307         .quad   0xc6e00bf33da88fc2,0xd5a79147930aa725
308         .quad   0x06ca6351e003826f,0x142929670a0e6e70
309         .quad   0x27b70a8546d22ffc,0x2e1b21385c26c926
310         .quad   0x4d2c6dfc5ac42aed,0x53380d139d95b3df
311         .quad   0x650a73548baf63de,0x766a0abb3c77b2a8
312         .quad   0x81c2c92e47edaee6,0x92722c851482353b
313         .quad   0xa2bfe8a14cf10364,0xa81a664bbc423001
314         .quad   0xc24b8b70d0f89791,0xc76c51a30654be30
315         .quad   0xd192e819d6ef5218,0xd69906245565a910
316         .quad   0xf40e35855771202a,0x106aa07032bbd1b8
317         .quad   0x19a4c116b8d2d0c8,0x1e376c085141ab53
318         .quad   0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
319         .quad   0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
320         .quad   0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
321         .quad   0x748f82ee5defb2fc,0x78a5636f43172f60
322         .quad   0x84c87814a1f0ab72,0x8cc702081a6439ec
323         .quad   0x90befffa23631e28,0xa4506cebde82bde9
324         .quad   0xbef9a3f7b2c67915,0xc67178f2e372532b
325         .quad   0xca273eceea26619c,0xd186b8c721c0c207
326         .quad   0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
327         .quad   0x06f067aa72176fba,0x0a637dc5a2c898a6
328         .quad   0x113f9804bef90dae,0x1b710b35131c471b
329         .quad   0x28db77f523047d84,0x32caab7b40c72493
330         .quad   0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
331         .quad   0x4cc5d4becb3e42b6,0x597f299cfc657e2a
332         .quad   0x5fcb6fab3ad6faec,0x6c44198c4a475817
333         .quad   0       // terminator
334 ___
335 $code.=<<___ if ($SZ==4);
336         .long   0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
337         .long   0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
338         .long   0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
339         .long   0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
340         .long   0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
341         .long   0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
342         .long   0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
343         .long   0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
344         .long   0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
345         .long   0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
346         .long   0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
347         .long   0xd192e819,0xd6990624,0xf40e3585,0x106aa070
348         .long   0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
349         .long   0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
350         .long   0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
351         .long   0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
352         .long   0       //terminator
353 ___
354 $code.=<<___;
355 .size   .LK$BITS,.-.LK$BITS
356 #ifndef __KERNEL__
357 .align  3
358 .LOPENSSL_armcap_P:
359 # ifdef __ILP32__
360         .long   OPENSSL_armcap_P-.
361 # else
362         .quad   OPENSSL_armcap_P-.
363 # endif
364 #endif
365 .asciz  "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
366 .align  2
367 ___
368
369 if ($SZ==4) {
370 my $Ktbl="x3";
371
372 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
373 my @MSG=map("v$_.16b",(4..7));
374 my ($W0,$W1)=("v16.4s","v17.4s");
375 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
376
377 $code.=<<___;
378 #ifndef __KERNEL__
379 .type   sha256_block_armv8,%function
380 .align  6
381 sha256_block_armv8:
382 .Lv8_entry:
383         stp             x29,x30,[sp,#-16]!
384         add             x29,sp,#0
385
386         ld1.32          {$ABCD,$EFGH},[$ctx]
387         adr             $Ktbl,.LK256
388
389 .Loop_hw:
390         ld1             {@MSG[0]-@MSG[3]},[$inp],#64
391         sub             $num,$num,#1
392         ld1.32          {$W0},[$Ktbl],#16
393         rev32           @MSG[0],@MSG[0]
394         rev32           @MSG[1],@MSG[1]
395         rev32           @MSG[2],@MSG[2]
396         rev32           @MSG[3],@MSG[3]
397         orr             $ABCD_SAVE,$ABCD,$ABCD          // offload
398         orr             $EFGH_SAVE,$EFGH,$EFGH
399 ___
400 for($i=0;$i<12;$i++) {
401 $code.=<<___;
402         ld1.32          {$W1},[$Ktbl],#16
403         add.i32         $W0,$W0,@MSG[0]
404         sha256su0       @MSG[0],@MSG[1]
405         orr             $abcd,$ABCD,$ABCD
406         sha256h         $ABCD,$EFGH,$W0
407         sha256h2        $EFGH,$abcd,$W0
408         sha256su1       @MSG[0],@MSG[2],@MSG[3]
409 ___
410         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
411 }
412 $code.=<<___;
413         ld1.32          {$W1},[$Ktbl],#16
414         add.i32         $W0,$W0,@MSG[0]
415         orr             $abcd,$ABCD,$ABCD
416         sha256h         $ABCD,$EFGH,$W0
417         sha256h2        $EFGH,$abcd,$W0
418
419         ld1.32          {$W0},[$Ktbl],#16
420         add.i32         $W1,$W1,@MSG[1]
421         orr             $abcd,$ABCD,$ABCD
422         sha256h         $ABCD,$EFGH,$W1
423         sha256h2        $EFGH,$abcd,$W1
424
425         ld1.32          {$W1},[$Ktbl]
426         add.i32         $W0,$W0,@MSG[2]
427         sub             $Ktbl,$Ktbl,#$rounds*$SZ-16     // rewind
428         orr             $abcd,$ABCD,$ABCD
429         sha256h         $ABCD,$EFGH,$W0
430         sha256h2        $EFGH,$abcd,$W0
431
432         add.i32         $W1,$W1,@MSG[3]
433         orr             $abcd,$ABCD,$ABCD
434         sha256h         $ABCD,$EFGH,$W1
435         sha256h2        $EFGH,$abcd,$W1
436
437         add.i32         $ABCD,$ABCD,$ABCD_SAVE
438         add.i32         $EFGH,$EFGH,$EFGH_SAVE
439
440         cbnz            $num,.Loop_hw
441
442         st1.32          {$ABCD,$EFGH},[$ctx]
443
444         ldr             x29,[sp],#16
445         ret
446 .size   sha256_block_armv8,.-sha256_block_armv8
447 #endif
448 ___
449 }
450
451 if ($SZ==4) {   ######################################### NEON stuff #
452 # You'll surely note a lot of similarities with sha256-armv4 module,
453 # and of course it's not a coincidence. sha256-armv4 was used as
454 # initial template, but was adapted for ARMv8 instruction set and
455 # extensively re-tuned for all-round performance.
456
457 my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
458 my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
459 my $Ktbl="x16";
460 my $Xfer="x17";
461 my @X = map("q$_",(0..3));
462 my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
463 my $j=0;
464
465 sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
466 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
467   my $arg = pop;
468     $arg = "#$arg" if ($arg*1 eq $arg);
469     $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
470 }
471
472 sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
473 sub Dlo     { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
474 sub Dhi     { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
475
476 sub Xupdate()
477 { use integer;
478   my $body = shift;
479   my @insns = (&$body,&$body,&$body,&$body);
480   my ($a,$b,$c,$d,$e,$f,$g,$h);
481
482         &ext_8          ($T0,@X[0],@X[1],4);    # X[1..4]
483          eval(shift(@insns));
484          eval(shift(@insns));
485          eval(shift(@insns));
486         &ext_8          ($T3,@X[2],@X[3],4);    # X[9..12]
487          eval(shift(@insns));
488          eval(shift(@insns));
489         &mov            (&Dscalar($T7),&Dhi(@X[3]));    # X[14..15]
490          eval(shift(@insns));
491          eval(shift(@insns));
492         &ushr_32        ($T2,$T0,$sigma0[0]);
493          eval(shift(@insns));
494         &ushr_32        ($T1,$T0,$sigma0[2]);
495          eval(shift(@insns));
496         &add_32         (@X[0],@X[0],$T3);      # X[0..3] += X[9..12]
497          eval(shift(@insns));
498         &sli_32         ($T2,$T0,32-$sigma0[0]);
499          eval(shift(@insns));
500          eval(shift(@insns));
501         &ushr_32        ($T3,$T0,$sigma0[1]);
502          eval(shift(@insns));
503          eval(shift(@insns));
504         &eor_8          ($T1,$T1,$T2);
505          eval(shift(@insns));
506          eval(shift(@insns));
507         &sli_32         ($T3,$T0,32-$sigma0[1]);
508          eval(shift(@insns));
509          eval(shift(@insns));
510           &ushr_32      ($T4,$T7,$sigma1[0]);
511          eval(shift(@insns));
512          eval(shift(@insns));
513         &eor_8          ($T1,$T1,$T3);          # sigma0(X[1..4])
514          eval(shift(@insns));
515          eval(shift(@insns));
516           &sli_32       ($T4,$T7,32-$sigma1[0]);
517          eval(shift(@insns));
518          eval(shift(@insns));
519           &ushr_32      ($T5,$T7,$sigma1[2]);
520          eval(shift(@insns));
521          eval(shift(@insns));
522           &ushr_32      ($T3,$T7,$sigma1[1]);
523          eval(shift(@insns));
524          eval(shift(@insns));
525         &add_32         (@X[0],@X[0],$T1);      # X[0..3] += sigma0(X[1..4])
526          eval(shift(@insns));
527          eval(shift(@insns));
528           &sli_u32      ($T3,$T7,32-$sigma1[1]);
529          eval(shift(@insns));
530          eval(shift(@insns));
531           &eor_8        ($T5,$T5,$T4);
532          eval(shift(@insns));
533          eval(shift(@insns));
534          eval(shift(@insns));
535           &eor_8        ($T5,$T5,$T3);          # sigma1(X[14..15])
536          eval(shift(@insns));
537          eval(shift(@insns));
538          eval(shift(@insns));
539         &add_32         (@X[0],@X[0],$T5);      # X[0..1] += sigma1(X[14..15])
540          eval(shift(@insns));
541          eval(shift(@insns));
542          eval(shift(@insns));
543           &ushr_32      ($T6,@X[0],$sigma1[0]);
544          eval(shift(@insns));
545           &ushr_32      ($T7,@X[0],$sigma1[2]);
546          eval(shift(@insns));
547          eval(shift(@insns));
548           &sli_32       ($T6,@X[0],32-$sigma1[0]);
549          eval(shift(@insns));
550           &ushr_32      ($T5,@X[0],$sigma1[1]);
551          eval(shift(@insns));
552          eval(shift(@insns));
553           &eor_8        ($T7,$T7,$T6);
554          eval(shift(@insns));
555          eval(shift(@insns));
556           &sli_32       ($T5,@X[0],32-$sigma1[1]);
557          eval(shift(@insns));
558          eval(shift(@insns));
559         &ld1_32         ("{$T0}","[$Ktbl], #16");
560          eval(shift(@insns));
561           &eor_8        ($T7,$T7,$T5);          # sigma1(X[16..17])
562          eval(shift(@insns));
563          eval(shift(@insns));
564         &eor_8          ($T5,$T5,$T5);
565          eval(shift(@insns));
566          eval(shift(@insns));
567         &mov            (&Dhi($T5), &Dlo($T7));
568          eval(shift(@insns));
569          eval(shift(@insns));
570          eval(shift(@insns));
571         &add_32         (@X[0],@X[0],$T5);      # X[2..3] += sigma1(X[16..17])
572          eval(shift(@insns));
573          eval(shift(@insns));
574          eval(shift(@insns));
575         &add_32         ($T0,$T0,@X[0]);
576          while($#insns>=1) { eval(shift(@insns)); }
577         &st1_32         ("{$T0}","[$Xfer], #16");
578          eval(shift(@insns));
579
580         push(@X,shift(@X));             # "rotate" X[]
581 }
582
583 sub Xpreload()
584 { use integer;
585   my $body = shift;
586   my @insns = (&$body,&$body,&$body,&$body);
587   my ($a,$b,$c,$d,$e,$f,$g,$h);
588
589          eval(shift(@insns));
590          eval(shift(@insns));
591         &ld1_8          ("{@X[0]}","[$inp],#16");
592          eval(shift(@insns));
593          eval(shift(@insns));
594         &ld1_32         ("{$T0}","[$Ktbl],#16");
595          eval(shift(@insns));
596          eval(shift(@insns));
597          eval(shift(@insns));
598          eval(shift(@insns));
599         &rev32          (@X[0],@X[0]);
600          eval(shift(@insns));
601          eval(shift(@insns));
602          eval(shift(@insns));
603          eval(shift(@insns));
604         &add_32         ($T0,$T0,@X[0]);
605          foreach (@insns) { eval; }     # remaining instructions
606         &st1_32         ("{$T0}","[$Xfer], #16");
607
608         push(@X,shift(@X));             # "rotate" X[]
609 }
610
611 sub body_00_15 () {
612         (
613         '($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
614         '&add   ($h,$h,$t1)',                   # h+=X[i]+K[i]
615         '&add   ($a,$a,$t4);'.                  # h+=Sigma0(a) from the past
616         '&and   ($t1,$f,$e)',
617         '&bic   ($t4,$g,$e)',
618         '&eor   ($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
619         '&add   ($a,$a,$t2)',                   # h+=Maj(a,b,c) from the past
620         '&orr   ($t1,$t1,$t4)',                 # Ch(e,f,g)
621         '&eor   ($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))',   # Sigma1(e)
622         '&eor   ($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
623         '&add   ($h,$h,$t1)',                   # h+=Ch(e,f,g)
624         '&ror   ($t0,$t0,"#$Sigma1[0]")',
625         '&eor   ($t2,$a,$b)',                   # a^b, b^c in next round
626         '&eor   ($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))',   # Sigma0(a)
627         '&add   ($h,$h,$t0)',                   # h+=Sigma1(e)
628         '&ldr   ($t1,sprintf "[sp,#%d]",4*(($j+1)&15))  if (($j&15)!=15);'.
629         '&ldr   ($t1,"[$Ktbl]")                         if ($j==15);'.
630         '&and   ($t3,$t3,$t2)',                 # (b^c)&=(a^b)
631         '&ror   ($t4,$t4,"#$Sigma0[0]")',
632         '&add   ($d,$d,$h)',                    # d+=h
633         '&eor   ($t3,$t3,$b)',                  # Maj(a,b,c)
634         '$j++;  unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
635         )
636 }
637
638 $code.=<<___;
639 #ifdef  __KERNEL__
640 .globl  sha256_block_neon
641 #endif
642 .type   sha256_block_neon,%function
643 .align  4
644 sha256_block_neon:
645 .Lneon_entry:
646         stp     x29, x30, [sp, #-16]!
647         mov     x29, sp
648         sub     sp,sp,#16*4
649
650         adr     $Ktbl,.LK256
651         add     $num,$inp,$num,lsl#6    // len to point at the end of inp
652
653         ld1.8   {@X[0]},[$inp], #16
654         ld1.8   {@X[1]},[$inp], #16
655         ld1.8   {@X[2]},[$inp], #16
656         ld1.8   {@X[3]},[$inp], #16
657         ld1.32  {$T0},[$Ktbl], #16
658         ld1.32  {$T1},[$Ktbl], #16
659         ld1.32  {$T2},[$Ktbl], #16
660         ld1.32  {$T3},[$Ktbl], #16
661         rev32   @X[0],@X[0]             // yes, even on
662         rev32   @X[1],@X[1]             // big-endian
663         rev32   @X[2],@X[2]
664         rev32   @X[3],@X[3]
665         mov     $Xfer,sp
666         add.32  $T0,$T0,@X[0]
667         add.32  $T1,$T1,@X[1]
668         add.32  $T2,$T2,@X[2]
669         st1.32  {$T0-$T1},[$Xfer], #32
670         add.32  $T3,$T3,@X[3]
671         st1.32  {$T2-$T3},[$Xfer]
672         sub     $Xfer,$Xfer,#32
673
674         ldp     $A,$B,[$ctx]
675         ldp     $C,$D,[$ctx,#8]
676         ldp     $E,$F,[$ctx,#16]
677         ldp     $G,$H,[$ctx,#24]
678         ldr     $t1,[sp,#0]
679         mov     $t2,wzr
680         eor     $t3,$B,$C
681         mov     $t4,wzr
682         b       .L_00_48
683
684 .align  4
685 .L_00_48:
686 ___
687         &Xupdate(\&body_00_15);
688         &Xupdate(\&body_00_15);
689         &Xupdate(\&body_00_15);
690         &Xupdate(\&body_00_15);
691 $code.=<<___;
692         cmp     $t1,#0                          // check for K256 terminator
693         ldr     $t1,[sp,#0]
694         sub     $Xfer,$Xfer,#64
695         bne     .L_00_48
696
697         sub     $Ktbl,$Ktbl,#256                // rewind $Ktbl
698         cmp     $inp,$num
699         mov     $Xfer, #64
700         csel    $Xfer, $Xfer, xzr, eq
701         sub     $inp,$inp,$Xfer                 // avoid SEGV
702         mov     $Xfer,sp
703 ___
704         &Xpreload(\&body_00_15);
705         &Xpreload(\&body_00_15);
706         &Xpreload(\&body_00_15);
707         &Xpreload(\&body_00_15);
708 $code.=<<___;
709         add     $A,$A,$t4                       // h+=Sigma0(a) from the past
710         ldp     $t0,$t1,[$ctx,#0]
711         add     $A,$A,$t2                       // h+=Maj(a,b,c) from the past
712         ldp     $t2,$t3,[$ctx,#8]
713         add     $A,$A,$t0                       // accumulate
714         add     $B,$B,$t1
715         ldp     $t0,$t1,[$ctx,#16]
716         add     $C,$C,$t2
717         add     $D,$D,$t3
718         ldp     $t2,$t3,[$ctx,#24]
719         add     $E,$E,$t0
720         add     $F,$F,$t1
721          ldr    $t1,[sp,#0]
722         stp     $A,$B,[$ctx,#0]
723         add     $G,$G,$t2
724          mov    $t2,wzr
725         stp     $C,$D,[$ctx,#8]
726         add     $H,$H,$t3
727         stp     $E,$F,[$ctx,#16]
728          eor    $t3,$B,$C
729         stp     $G,$H,[$ctx,#24]
730          mov    $t4,wzr
731          mov    $Xfer,sp
732         b.ne    .L_00_48
733
734         ldr     x29,[x29]
735         add     sp,sp,#16*4+16
736         ret
737 .size   sha256_block_neon,.-sha256_block_neon
738 ___
739 }
740
741 if ($SZ==8) {
742 my $Ktbl="x3";
743
744 my @H = map("v$_.16b",(0..4));
745 my ($fg,$de,$m9_10)=map("v$_.16b",(5..7));
746 my @MSG=map("v$_.16b",(16..23));
747 my ($W0,$W1)=("v24.2d","v25.2d");
748 my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29));
749
750 $code.=<<___;
751 #ifndef __KERNEL__
752 .type   sha512_block_armv8,%function
753 .align  6
754 sha512_block_armv8:
755 .Lv8_entry:
756         stp             x29,x30,[sp,#-16]!
757         add             x29,sp,#0
758
759         ld1             {@MSG[0]-@MSG[3]},[$inp],#64    // load input
760         ld1             {@MSG[4]-@MSG[7]},[$inp],#64
761
762         ld1.64          {@H[0]-@H[3]},[$ctx]            // load context
763         adr             $Ktbl,.LK512
764
765         rev64           @MSG[0],@MSG[0]
766         rev64           @MSG[1],@MSG[1]
767         rev64           @MSG[2],@MSG[2]
768         rev64           @MSG[3],@MSG[3]
769         rev64           @MSG[4],@MSG[4]
770         rev64           @MSG[5],@MSG[5]
771         rev64           @MSG[6],@MSG[6]
772         rev64           @MSG[7],@MSG[7]
773         b               .Loop_hw
774
775 .align  4
776 .Loop_hw:
777         ld1.64          {$W0},[$Ktbl],#16
778         subs            $num,$num,#1
779         sub             x4,$inp,#128
780         orr             $AB,@H[0],@H[0]                 // offload
781         orr             $CD,@H[1],@H[1]
782         orr             $EF,@H[2],@H[2]
783         orr             $GH,@H[3],@H[3]
784         csel            $inp,$inp,x4,ne                 // conditional rewind
785 ___
786 for($i=0;$i<32;$i++) {
787 $code.=<<___;
788         add.i64         $W0,$W0,@MSG[0]
789         ld1.64          {$W1},[$Ktbl],#16
790         ext             $W0,$W0,$W0,#8
791         ext             $fg,@H[2],@H[3],#8
792         ext             $de,@H[1],@H[2],#8
793         add.i64         @H[3],@H[3],$W0                 // "T1 + H + K512[i]"
794          sha512su0      @MSG[0],@MSG[1]
795          ext            $m9_10,@MSG[4],@MSG[5],#8
796         sha512h         @H[3],$fg,$de
797          sha512su1      @MSG[0],@MSG[7],$m9_10
798         add.i64         @H[4],@H[1],@H[3]               // "D + T1"
799         sha512h2        @H[3],$H[1],@H[0]
800 ___
801         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
802         @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
803 }
804 for(;$i<40;$i++) {
805 $code.=<<___    if ($i<39);
806         ld1.64          {$W1},[$Ktbl],#16
807 ___
808 $code.=<<___    if ($i==39);
809         sub             $Ktbl,$Ktbl,#$rounds*$SZ        // rewind
810 ___
811 $code.=<<___;
812         add.i64         $W0,$W0,@MSG[0]
813          ld1            {@MSG[0]},[$inp],#16            // load next input
814         ext             $W0,$W0,$W0,#8
815         ext             $fg,@H[2],@H[3],#8
816         ext             $de,@H[1],@H[2],#8
817         add.i64         @H[3],@H[3],$W0                 // "T1 + H + K512[i]"
818         sha512h         @H[3],$fg,$de
819          rev64          @MSG[0],@MSG[0]
820         add.i64         @H[4],@H[1],@H[3]               // "D + T1"
821         sha512h2        @H[3],$H[1],@H[0]
822 ___
823         ($W0,$W1)=($W1,$W0);    push(@MSG,shift(@MSG));
824         @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
825 }
826 $code.=<<___;
827         add.i64         @H[0],@H[0],$AB                 // accumulate
828         add.i64         @H[1],@H[1],$CD
829         add.i64         @H[2],@H[2],$EF
830         add.i64         @H[3],@H[3],$GH
831
832         cbnz            $num,.Loop_hw
833
834         st1.64          {@H[0]-@H[3]},[$ctx]            // store context
835
836         ldr             x29,[sp],#16
837         ret
838 .size   sha512_block_armv8,.-sha512_block_armv8
839 #endif
840 ___
841 }
842
843 $code.=<<___;
844 #ifndef __KERNEL__
845 .comm   OPENSSL_armcap_P,4,4
846 #endif
847 ___
848
849 {   my  %opcode = (
850         "sha256h"       => 0x5e004000,  "sha256h2"      => 0x5e005000,
851         "sha256su0"     => 0x5e282800,  "sha256su1"     => 0x5e006000   );
852
853     sub unsha256 {
854         my ($mnemonic,$arg)=@_;
855
856         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
857         &&
858         sprintf ".inst\t0x%08x\t//%s %s",
859                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
860                         $mnemonic,$arg;
861     }
862 }
863
864 {   my  %opcode = (
865         "sha512h"       => 0xce608000,  "sha512h2"      => 0xce608400,
866         "sha512su0"     => 0xcec08000,  "sha512su1"     => 0xce608800   );
867
868     sub unsha512 {
869         my ($mnemonic,$arg)=@_;
870
871         $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
872         &&
873         sprintf ".inst\t0x%08x\t//%s %s",
874                         $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
875                         $mnemonic,$arg;
876     }
877 }
878
879 open SELF,$0;
880 while(<SELF>) {
881         next if (/^#!/);
882         last if (!s/^#/\/\// and !/^$/);
883         print;
884 }
885 close SELF;
886
887 foreach(split("\n",$code)) {
888
889         s/\`([^\`]*)\`/eval($1)/ge;
890
891         s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge   or
892         s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
893
894         s/\bq([0-9]+)\b/v$1.16b/g;              # old->new registers
895
896         s/\.[ui]?8(\s)/$1/;
897         s/\.\w?64\b//           and s/\.16b/\.2d/g      or
898         s/\.\w?32\b//           and s/\.16b/\.4s/g;
899         m/\bext\b/              and s/\.2d/\.16b/g      or
900         m/(ld|st)1[^\[]+\[0\]/  and s/\.4s/\.s/g;
901
902         print $_,"\n";
903 }
904
905 close STDOUT;