2 # Copyright 2014-2020 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the Apache License 2.0 (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
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/.
15 # Permission to use under GPLv2 terms is granted.
16 # ====================================================================
18 # SHA256/512 for ARMv8.
20 # Performance in cycles per processed byte and improvement coefficient
21 # over code generated with "default" compiler:
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 # ThunderX2 2.54 13.2 (+40%) 8.40 (+18%)
33 # (*) Software SHA256 results are of lesser relevance, presented
34 # mostly for informational purposes.
35 # (**) The result is a trade-off: it's possible to improve it by
36 # 10% (or by 1 cycle per round), but at the cost of 20% loss
37 # on Cortex-A53 (or by 4 cycles per round).
38 # (***) Super-impressive coefficients over gcc-generated code are
39 # indication of some compiler "pathology", most notably code
40 # generated with -mgeneral-regs-only is significantly faster
41 # and the gap is only 40-90%.
45 # Originally it was reckoned that it makes no sense to implement NEON
46 # version of SHA256 for 64-bit processors. This is because performance
47 # improvement on most wide-spread Cortex-A5x processors was observed
48 # to be marginal, same on Cortex-A53 and ~10% on A57. But then it was
49 # observed that 32-bit NEON SHA256 performs significantly better than
50 # 64-bit scalar version on *some* of the more recent processors. As
51 # result 64-bit NEON version of SHA256 was added to provide best
52 # all-round performance. For example it executes ~30% faster on X-Gene
53 # and Mongoose. [For reference, NEON version of SHA512 is bound to
54 # deliver much less improvement, likely *negative* on Cortex-A5x.
55 # Which is why NEON support is limited to SHA256.]
57 # $output is the last argument if it looks like a file (it has an extension)
58 # $flavour is the first argument if it doesn't look like a file
59 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
60 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
62 if ($flavour && $flavour ne "void") {
63 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
64 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
65 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
66 die "can't locate arm-xlate.pl";
68 open OUT,"| \"$^X\" $xlate $flavour \"$output\""
69 or die "can't call $xlate: $!";
72 $output and open STDOUT,">$output";
75 if ($output =~ /512/) {
95 $func="sha${BITS}_block_data_order";
97 ($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
99 @X=map("$reg_t$_",(3..15,0..2));
100 @V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
101 ($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
104 my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
106 my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
107 $T0=@X[$i+3] if ($i<11);
109 $code.=<<___ if ($i<16);
110 #ifndef __AARCH64EB__
111 rev @X[$i],@X[$i] // $i
114 $code.=<<___ if ($i<13 && ($i&1));
115 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ
117 $code.=<<___ if ($i==13);
118 ldp @X[14],@X[15],[$inp]
120 $code.=<<___ if ($i>=14);
121 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
123 $code.=<<___ if ($i>0 && $i<16);
124 add $a,$a,$t1 // h+=Sigma0(a)
126 $code.=<<___ if ($i>=11);
127 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
129 # While ARMv8 specifies merged rotate-n-logical operation such as
130 # 'eor x,y,z,ror#n', it was found to negatively affect performance
131 # on Apple A7. The reason seems to be that it requires even 'y' to
132 # be available earlier. This means that such merged instruction is
133 # not necessarily best choice on critical path... On the other hand
134 # Cortex-A5x handles merged instructions much better than disjoint
135 # rotate and logical... See (**) footnote above.
136 $code.=<<___ if ($i<15);
137 ror $t0,$e,#$Sigma1[0]
138 add $h,$h,$t2 // h+=K[i]
139 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
142 add $h,$h,@X[$i&15] // h+=X[i]
143 orr $t1,$t1,$t2 // Ch(e,f,g)
144 eor $t2,$a,$b // a^b, b^c in next round
145 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e)
146 ror $T0,$a,#$Sigma0[0]
147 add $h,$h,$t1 // h+=Ch(e,f,g)
148 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
149 add $h,$h,$t0 // h+=Sigma1(e)
150 and $t3,$t3,$t2 // (b^c)&=(a^b)
152 eor $t3,$t3,$b // Maj(a,b,c)
153 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a)
154 add $h,$h,$t3 // h+=Maj(a,b,c)
155 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
156 //add $h,$h,$t1 // h+=Sigma0(a)
158 $code.=<<___ if ($i>=15);
159 ror $t0,$e,#$Sigma1[0]
160 add $h,$h,$t2 // h+=K[i]
161 ror $T1,@X[($j+1)&15],#$sigma0[0]
163 ror $T2,@X[($j+14)&15],#$sigma1[0]
165 ror $T0,$a,#$Sigma0[0]
166 add $h,$h,@X[$i&15] // h+=X[i]
167 eor $t0,$t0,$e,ror#$Sigma1[1]
168 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
169 orr $t1,$t1,$t2 // Ch(e,f,g)
170 eor $t2,$a,$b // a^b, b^c in next round
171 eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e)
172 eor $T0,$T0,$a,ror#$Sigma0[1]
173 add $h,$h,$t1 // h+=Ch(e,f,g)
174 and $t3,$t3,$t2 // (b^c)&=(a^b)
175 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
176 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1])
177 add $h,$h,$t0 // h+=Sigma1(e)
178 eor $t3,$t3,$b // Maj(a,b,c)
179 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a)
180 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14])
181 add @X[$j],@X[$j],@X[($j+9)&15]
183 add $h,$h,$t3 // h+=Maj(a,b,c)
184 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round
185 add @X[$j],@X[$j],$T1
186 add $h,$h,$t1 // h+=Sigma0(a)
187 add @X[$j],@X[$j],$T2
193 #include "arm_arch.h"
195 .extern OPENSSL_armcap_P
196 .hidden OPENSSL_armcap_P
202 .type $func,%function
205 AARCH64_VALID_CALL_TARGET
207 adrp x16,OPENSSL_armcap_P
208 ldr w16,[x16,#:lo12:OPENSSL_armcap_P]
210 $code.=<<___ if ($SZ==4);
211 tst w16,#ARMV8_SHA256
216 $code.=<<___ if ($SZ==8);
217 tst w16,#ARMV8_SHA512
222 AARCH64_SIGN_LINK_REGISTER
223 stp x29,x30,[sp,#-128]!
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]
239 stp $ctx,$num,[x29,#96]
242 ldp @X[0],@X[1],[$inp],#2*$SZ
243 ldr $t2,[$Ktbl],#$SZ // *K++
244 eor $t3,$B,$C // magic seed
247 for ($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)); }
253 ldp $ctx,$num,[x29,#96]
255 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind
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*$SZ]
262 ldp @X[6],@X[7],[$ctx,#6*$SZ]
269 stp $C,$D,[$ctx,#2*$SZ]
273 stp $E,$F,[$ctx,#4*$SZ]
274 stp $G,$H,[$ctx,#6*$SZ]
277 ldp x19,x20,[x29,#16]
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 AARCH64_VALIDATE_LINK_REGISTER
289 .type .LK$BITS,%object
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
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
355 .size .LK$BITS,.-.LK$BITS
356 .asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
363 my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
364 my @MSG=map("v$_.16b",(4..7));
365 my ($W0,$W1)=("v16.4s","v17.4s");
366 my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
370 .type sha256_block_armv8,%function
374 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later.
375 stp x29,x30,[sp,#-16]!
378 ld1.32 {$ABCD,$EFGH},[$ctx]
382 ld1 {@MSG[0]-@MSG[3]},[$inp],#64
384 ld1.32 {$W0},[$Ktbl],#16
385 rev32 @MSG[0],@MSG[0]
386 rev32 @MSG[1],@MSG[1]
387 rev32 @MSG[2],@MSG[2]
388 rev32 @MSG[3],@MSG[3]
389 orr $ABCD_SAVE,$ABCD,$ABCD // offload
390 orr $EFGH_SAVE,$EFGH,$EFGH
392 for($i=0;$i<12;$i++) {
394 ld1.32 {$W1},[$Ktbl],#16
395 add.i32 $W0,$W0,@MSG[0]
396 sha256su0 @MSG[0],@MSG[1]
397 orr $abcd,$ABCD,$ABCD
398 sha256h $ABCD,$EFGH,$W0
399 sha256h2 $EFGH,$abcd,$W0
400 sha256su1 @MSG[0],@MSG[2],@MSG[3]
402 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG));
405 ld1.32 {$W1},[$Ktbl],#16
406 add.i32 $W0,$W0,@MSG[0]
407 orr $abcd,$ABCD,$ABCD
408 sha256h $ABCD,$EFGH,$W0
409 sha256h2 $EFGH,$abcd,$W0
411 ld1.32 {$W0},[$Ktbl],#16
412 add.i32 $W1,$W1,@MSG[1]
413 orr $abcd,$ABCD,$ABCD
414 sha256h $ABCD,$EFGH,$W1
415 sha256h2 $EFGH,$abcd,$W1
418 add.i32 $W0,$W0,@MSG[2]
419 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind
420 orr $abcd,$ABCD,$ABCD
421 sha256h $ABCD,$EFGH,$W0
422 sha256h2 $EFGH,$abcd,$W0
424 add.i32 $W1,$W1,@MSG[3]
425 orr $abcd,$ABCD,$ABCD
426 sha256h $ABCD,$EFGH,$W1
427 sha256h2 $EFGH,$abcd,$W1
429 add.i32 $ABCD,$ABCD,$ABCD_SAVE
430 add.i32 $EFGH,$EFGH,$EFGH_SAVE
434 st1.32 {$ABCD,$EFGH},[$ctx]
438 .size sha256_block_armv8,.-sha256_block_armv8
443 if ($SZ==4) { ######################################### NEON stuff #
444 # You'll surely note a lot of similarities with sha256-armv4 module,
445 # and of course it's not a coincidence. sha256-armv4 was used as
446 # initial template, but was adapted for ARMv8 instruction set and
447 # extensively re-tuned for all-round performance.
449 my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
450 my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
453 my @X = map("q$_",(0..3));
454 my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
457 sub AUTOLOAD() # thunk [simplified] x86-style perlasm
458 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
460 $arg = "#$arg" if ($arg*1 eq $arg);
461 $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
464 sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
465 sub Dlo { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
466 sub Dhi { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
471 my @insns = (&$body,&$body,&$body,&$body);
472 my ($a,$b,$c,$d,$e,$f,$g,$h);
474 &ext_8 ($T0,@X[0],@X[1],4); # X[1..4]
478 &ext_8 ($T3,@X[2],@X[3],4); # X[9..12]
481 &mov (&Dscalar($T7),&Dhi(@X[3])); # X[14..15]
484 &ushr_32 ($T2,$T0,$sigma0[0]);
486 &ushr_32 ($T1,$T0,$sigma0[2]);
488 &add_32 (@X[0],@X[0],$T3); # X[0..3] += X[9..12]
490 &sli_32 ($T2,$T0,32-$sigma0[0]);
493 &ushr_32 ($T3,$T0,$sigma0[1]);
496 &eor_8 ($T1,$T1,$T2);
499 &sli_32 ($T3,$T0,32-$sigma0[1]);
502 &ushr_32 ($T4,$T7,$sigma1[0]);
505 &eor_8 ($T1,$T1,$T3); # sigma0(X[1..4])
508 &sli_32 ($T4,$T7,32-$sigma1[0]);
511 &ushr_32 ($T5,$T7,$sigma1[2]);
514 &ushr_32 ($T3,$T7,$sigma1[1]);
517 &add_32 (@X[0],@X[0],$T1); # X[0..3] += sigma0(X[1..4])
520 &sli_u32 ($T3,$T7,32-$sigma1[1]);
523 &eor_8 ($T5,$T5,$T4);
527 &eor_8 ($T5,$T5,$T3); # sigma1(X[14..15])
531 &add_32 (@X[0],@X[0],$T5); # X[0..1] += sigma1(X[14..15])
535 &ushr_32 ($T6,@X[0],$sigma1[0]);
537 &ushr_32 ($T7,@X[0],$sigma1[2]);
540 &sli_32 ($T6,@X[0],32-$sigma1[0]);
542 &ushr_32 ($T5,@X[0],$sigma1[1]);
545 &eor_8 ($T7,$T7,$T6);
548 &sli_32 ($T5,@X[0],32-$sigma1[1]);
551 &ld1_32 ("{$T0}","[$Ktbl], #16");
553 &eor_8 ($T7,$T7,$T5); # sigma1(X[16..17])
556 &eor_8 ($T5,$T5,$T5);
559 &mov (&Dhi($T5), &Dlo($T7));
563 &add_32 (@X[0],@X[0],$T5); # X[2..3] += sigma1(X[16..17])
567 &add_32 ($T0,$T0,@X[0]);
568 while($#insns>=1) { eval(shift(@insns)); }
569 &st1_32 ("{$T0}","[$Xfer], #16");
572 push(@X,shift(@X)); # "rotate" X[]
578 my @insns = (&$body,&$body,&$body,&$body);
579 my ($a,$b,$c,$d,$e,$f,$g,$h);
583 &ld1_8 ("{@X[0]}","[$inp],#16");
586 &ld1_32 ("{$T0}","[$Ktbl],#16");
591 &rev32 (@X[0],@X[0]);
596 &add_32 ($T0,$T0,@X[0]);
597 foreach (@insns) { eval; } # remaining instructions
598 &st1_32 ("{$T0}","[$Xfer], #16");
600 push(@X,shift(@X)); # "rotate" X[]
605 '($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
606 '&add ($h,$h,$t1)', # h+=X[i]+K[i]
607 '&add ($a,$a,$t4);'. # h+=Sigma0(a) from the past
610 '&eor ($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
611 '&add ($a,$a,$t2)', # h+=Maj(a,b,c) from the past
612 '&orr ($t1,$t1,$t4)', # Ch(e,f,g)
613 '&eor ($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))', # Sigma1(e)
614 '&eor ($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
615 '&add ($h,$h,$t1)', # h+=Ch(e,f,g)
616 '&ror ($t0,$t0,"#$Sigma1[0]")',
617 '&eor ($t2,$a,$b)', # a^b, b^c in next round
618 '&eor ($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))', # Sigma0(a)
619 '&add ($h,$h,$t0)', # h+=Sigma1(e)
620 '&ldr ($t1,sprintf "[sp,#%d]",4*(($j+1)&15)) if (($j&15)!=15);'.
621 '&ldr ($t1,"[$Ktbl]") if ($j==15);'.
622 '&and ($t3,$t3,$t2)', # (b^c)&=(a^b)
623 '&ror ($t4,$t4,"#$Sigma0[0]")',
624 '&add ($d,$d,$h)', # d+=h
625 '&eor ($t3,$t3,$b)', # Maj(a,b,c)
626 '$j++; unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
632 .globl sha256_block_neon
634 .type sha256_block_neon,%function
637 AARCH64_VALID_CALL_TARGET
639 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later
640 stp x29, x30, [sp, #-16]!
645 add $num,$inp,$num,lsl#6 // len to point at the end of inp
647 ld1.8 {@X[0]},[$inp], #16
648 ld1.8 {@X[1]},[$inp], #16
649 ld1.8 {@X[2]},[$inp], #16
650 ld1.8 {@X[3]},[$inp], #16
651 ld1.32 {$T0},[$Ktbl], #16
652 ld1.32 {$T1},[$Ktbl], #16
653 ld1.32 {$T2},[$Ktbl], #16
654 ld1.32 {$T3},[$Ktbl], #16
655 rev32 @X[0],@X[0] // yes, even on
656 rev32 @X[1],@X[1] // big-endian
663 st1.32 {$T0-$T1},[$Xfer], #32
665 st1.32 {$T2-$T3},[$Xfer]
681 &Xupdate(\&body_00_15);
682 &Xupdate(\&body_00_15);
683 &Xupdate(\&body_00_15);
684 &Xupdate(\&body_00_15);
686 cmp $t1,#0 // check for K256 terminator
691 sub $Ktbl,$Ktbl,#256 // rewind $Ktbl
694 csel $Xfer, $Xfer, xzr, eq
695 sub $inp,$inp,$Xfer // avoid SEGV
698 &Xpreload(\&body_00_15);
699 &Xpreload(\&body_00_15);
700 &Xpreload(\&body_00_15);
701 &Xpreload(\&body_00_15);
703 add $A,$A,$t4 // h+=Sigma0(a) from the past
704 ldp $t0,$t1,[$ctx,#0]
705 add $A,$A,$t2 // h+=Maj(a,b,c) from the past
706 ldp $t2,$t3,[$ctx,#8]
707 add $A,$A,$t0 // accumulate
709 ldp $t0,$t1,[$ctx,#16]
712 ldp $t2,$t3,[$ctx,#24]
731 .size sha256_block_neon,.-sha256_block_neon
738 my @H = map("v$_.16b",(0..4));
739 my ($fg,$de,$m9_10)=map("v$_.16b",(5..7));
740 my @MSG=map("v$_.16b",(16..23));
741 my ($W0,$W1)=("v24.2d","v25.2d");
742 my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29));
746 .type sha512_block_armv8,%function
750 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later
751 stp x29,x30,[sp,#-16]!
754 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 // load input
755 ld1 {@MSG[4]-@MSG[7]},[$inp],#64
757 ld1.64 {@H[0]-@H[3]},[$ctx] // load context
760 rev64 @MSG[0],@MSG[0]
761 rev64 @MSG[1],@MSG[1]
762 rev64 @MSG[2],@MSG[2]
763 rev64 @MSG[3],@MSG[3]
764 rev64 @MSG[4],@MSG[4]
765 rev64 @MSG[5],@MSG[5]
766 rev64 @MSG[6],@MSG[6]
767 rev64 @MSG[7],@MSG[7]
772 ld1.64 {$W0},[$Ktbl],#16
775 orr $AB,@H[0],@H[0] // offload
779 csel $inp,$inp,x4,ne // conditional rewind
781 for($i=0;$i<32;$i++) {
783 add.i64 $W0,$W0,@MSG[0]
784 ld1.64 {$W1},[$Ktbl],#16
786 ext $fg,@H[2],@H[3],#8
787 ext $de,@H[1],@H[2],#8
788 add.i64 @H[3],@H[3],$W0 // "T1 + H + K512[i]"
789 sha512su0 @MSG[0],@MSG[1]
790 ext $m9_10,@MSG[4],@MSG[5],#8
791 sha512h @H[3],$fg,$de
792 sha512su1 @MSG[0],@MSG[7],$m9_10
793 add.i64 @H[4],@H[1],@H[3] // "D + T1"
794 sha512h2 @H[3],$H[1],@H[0]
796 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG));
797 @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
800 $code.=<<___ if ($i<39);
801 ld1.64 {$W1},[$Ktbl],#16
803 $code.=<<___ if ($i==39);
804 sub $Ktbl,$Ktbl,#$rounds*$SZ // rewind
807 add.i64 $W0,$W0,@MSG[0]
808 ld1 {@MSG[0]},[$inp],#16 // load next input
810 ext $fg,@H[2],@H[3],#8
811 ext $de,@H[1],@H[2],#8
812 add.i64 @H[3],@H[3],$W0 // "T1 + H + K512[i]"
813 sha512h @H[3],$fg,$de
814 rev64 @MSG[0],@MSG[0]
815 add.i64 @H[4],@H[1],@H[3] // "D + T1"
816 sha512h2 @H[3],$H[1],@H[0]
818 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG));
819 @H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
822 add.i64 @H[0],@H[0],$AB // accumulate
823 add.i64 @H[1],@H[1],$CD
824 add.i64 @H[2],@H[2],$EF
825 add.i64 @H[3],@H[3],$GH
829 st1.64 {@H[0]-@H[3]},[$ctx] // store context
833 .size sha512_block_armv8,.-sha512_block_armv8
839 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000,
840 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 );
843 my ($mnemonic,$arg)=@_;
845 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
847 sprintf ".inst\t0x%08x\t//%s %s",
848 $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
854 "sha512h" => 0xce608000, "sha512h2" => 0xce608400,
855 "sha512su0" => 0xcec08000, "sha512su1" => 0xce608800 );
858 my ($mnemonic,$arg)=@_;
860 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
862 sprintf ".inst\t0x%08x\t//%s %s",
863 $opcode{$mnemonic}|$1|($2<<5)|($3<<16),
871 last if (!s/^#/\/\// and !/^$/);
876 foreach(split("\n",$code)) {
878 s/\`([^\`]*)\`/eval($1)/ge;
880 s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge or
881 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
883 s/\bq([0-9]+)\b/v$1.16b/g; # old->new registers
886 s/\.\w?64\b// and s/\.16b/\.2d/g or
887 s/\.\w?32\b// and s/\.16b/\.4s/g;
888 m/\bext\b/ and s/\.2d/\.16b/g or
889 m/(ld|st)1[^\[]+\[0\]/ and s/\.4s/\.s/g;
894 close STDOUT or die "error closing STDOUT: $!";