2 # Copyright 2016-2020 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
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 # ====================================================================
23 # Replace 3xNEON+1xIALU code path with 4+1. 4+1 is actually fastest
24 # option on most(*), but not all, processors, yet 6+2 is retained.
25 # This is because penalties are considered tolerable in comparison to
26 # improvement on processors where 6+2 helps. Most notably +37% on
27 # ThunderX2. It's server-oriented processor which will have to serve
28 # as many requests as possible. While others are mostly clients, when
29 # performance doesn't have to be absolute top-notch, just fast enough,
30 # as majority of time is spent "entertaining" relatively slow human.
32 # Performance in cycles per byte out of large buffer.
34 # IALU/gcc-4.9 4xNEON+1xIALU 6xNEON+2xIALU
36 # Apple A7 5.50/+49% 2.72 1.60
37 # Cortex-A53 8.40/+80% 4.06 4.45(*)
38 # Cortex-A57 8.06/+43% 4.15 4.40(*)
39 # Denver 4.50/+82% 2.30 2.70(*)
40 # X-Gene 9.50/+46% 8.20 8.90(*)
41 # Mongoose 8.00/+44% 2.74 3.12(*)
42 # Kryo 8.17/+50% 4.47 4.65(*)
43 # ThunderX2 7.22/+48% 5.64 4.10
45 # (*) slower than 4+1:-(
47 # $output is the last argument if it looks like a file (it has an extension)
48 # $flavour is the first argument if it doesn't look like a file
49 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
50 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
52 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
53 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
54 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
55 die "can't locate arm-xlate.pl";
57 open OUT,"| \"$^X\" $xlate $flavour \"$output\""
58 or die "can't call $xlate: $!";
61 sub AUTOLOAD() # thunk [simplified] x86-style perlasm
62 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
64 $arg = "#$arg" if ($arg*1 eq $arg);
65 $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
68 my ($out,$inp,$len,$key,$ctr) = map("x$_",(0..4));
70 my @x=map("x$_",(5..17,19..21));
71 my @d=map("x$_",(22..28,30));
74 my ($a0,$b0,$c0,$d0)=@_;
75 my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
76 my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
77 my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
80 "&add_32 (@x[$a0],@x[$a0],@x[$b0])",
81 "&add_32 (@x[$a1],@x[$a1],@x[$b1])",
82 "&add_32 (@x[$a2],@x[$a2],@x[$b2])",
83 "&add_32 (@x[$a3],@x[$a3],@x[$b3])",
84 "&eor_32 (@x[$d0],@x[$d0],@x[$a0])",
85 "&eor_32 (@x[$d1],@x[$d1],@x[$a1])",
86 "&eor_32 (@x[$d2],@x[$d2],@x[$a2])",
87 "&eor_32 (@x[$d3],@x[$d3],@x[$a3])",
88 "&ror_32 (@x[$d0],@x[$d0],16)",
89 "&ror_32 (@x[$d1],@x[$d1],16)",
90 "&ror_32 (@x[$d2],@x[$d2],16)",
91 "&ror_32 (@x[$d3],@x[$d3],16)",
93 "&add_32 (@x[$c0],@x[$c0],@x[$d0])",
94 "&add_32 (@x[$c1],@x[$c1],@x[$d1])",
95 "&add_32 (@x[$c2],@x[$c2],@x[$d2])",
96 "&add_32 (@x[$c3],@x[$c3],@x[$d3])",
97 "&eor_32 (@x[$b0],@x[$b0],@x[$c0])",
98 "&eor_32 (@x[$b1],@x[$b1],@x[$c1])",
99 "&eor_32 (@x[$b2],@x[$b2],@x[$c2])",
100 "&eor_32 (@x[$b3],@x[$b3],@x[$c3])",
101 "&ror_32 (@x[$b0],@x[$b0],20)",
102 "&ror_32 (@x[$b1],@x[$b1],20)",
103 "&ror_32 (@x[$b2],@x[$b2],20)",
104 "&ror_32 (@x[$b3],@x[$b3],20)",
106 "&add_32 (@x[$a0],@x[$a0],@x[$b0])",
107 "&add_32 (@x[$a1],@x[$a1],@x[$b1])",
108 "&add_32 (@x[$a2],@x[$a2],@x[$b2])",
109 "&add_32 (@x[$a3],@x[$a3],@x[$b3])",
110 "&eor_32 (@x[$d0],@x[$d0],@x[$a0])",
111 "&eor_32 (@x[$d1],@x[$d1],@x[$a1])",
112 "&eor_32 (@x[$d2],@x[$d2],@x[$a2])",
113 "&eor_32 (@x[$d3],@x[$d3],@x[$a3])",
114 "&ror_32 (@x[$d0],@x[$d0],24)",
115 "&ror_32 (@x[$d1],@x[$d1],24)",
116 "&ror_32 (@x[$d2],@x[$d2],24)",
117 "&ror_32 (@x[$d3],@x[$d3],24)",
119 "&add_32 (@x[$c0],@x[$c0],@x[$d0])",
120 "&add_32 (@x[$c1],@x[$c1],@x[$d1])",
121 "&add_32 (@x[$c2],@x[$c2],@x[$d2])",
122 "&add_32 (@x[$c3],@x[$c3],@x[$d3])",
123 "&eor_32 (@x[$b0],@x[$b0],@x[$c0])",
124 "&eor_32 (@x[$b1],@x[$b1],@x[$c1])",
125 "&eor_32 (@x[$b2],@x[$b2],@x[$c2])",
126 "&eor_32 (@x[$b3],@x[$b3],@x[$c3])",
127 "&ror_32 (@x[$b0],@x[$b0],25)",
128 "&ror_32 (@x[$b1],@x[$b1],25)",
129 "&ror_32 (@x[$b2],@x[$b2],25)",
130 "&ror_32 (@x[$b3],@x[$b3],25)"
136 # include "arm_arch.h"
137 .extern OPENSSL_armcap_P
144 .quad 0x3320646e61707865,0x6b20657479622d32 // endian-neutral
148 .long 0x02010003,0x06050407,0x0a09080b,0x0e0d0c0f
149 .asciz "ChaCha20 for ARMv8, CRYPTOGAMS by \@dot-asm"
151 .globl ChaCha20_ctr32
152 .type ChaCha20_ctr32,%function
160 adrp x17,OPENSSL_armcap_P
161 ldr w17,[x17,#:lo12:OPENSSL_armcap_P]
167 .inst 0xd503233f // paciasp
168 stp x29,x30,[sp,#-96]!
179 ldp @d[0],@d[1],[@x[0]] // load sigma
180 ldp @d[2],@d[3],[$key] // load key
181 ldp @d[4],@d[5],[$key,#16]
182 ldp @d[6],@d[7],[$ctr] // load counter
193 mov.32 @x[0],@d[0] // unpack key block
215 foreach (&ROUND(0, 4, 8,12)) { eval; }
216 foreach (&ROUND(0, 5,10,15)) { eval; }
220 add.32 @x[0],@x[0],@d[0] // accumulate key block
221 add @x[1],@x[1],@d[0],lsr#32
222 add.32 @x[2],@x[2],@d[1]
223 add @x[3],@x[3],@d[1],lsr#32
224 add.32 @x[4],@x[4],@d[2]
225 add @x[5],@x[5],@d[2],lsr#32
226 add.32 @x[6],@x[6],@d[3]
227 add @x[7],@x[7],@d[3],lsr#32
228 add.32 @x[8],@x[8],@d[4]
229 add @x[9],@x[9],@d[4],lsr#32
230 add.32 @x[10],@x[10],@d[5]
231 add @x[11],@x[11],@d[5],lsr#32
232 add.32 @x[12],@x[12],@d[6]
233 add @x[13],@x[13],@d[6],lsr#32
234 add.32 @x[14],@x[14],@d[7]
235 add @x[15],@x[15],@d[7],lsr#32
239 add @x[0],@x[0],@x[1],lsl#32 // pack
240 add @x[2],@x[2],@x[3],lsl#32
241 ldp @x[1],@x[3],[$inp,#0] // load input
242 add @x[4],@x[4],@x[5],lsl#32
243 add @x[6],@x[6],@x[7],lsl#32
244 ldp @x[5],@x[7],[$inp,#16]
245 add @x[8],@x[8],@x[9],lsl#32
246 add @x[10],@x[10],@x[11],lsl#32
247 ldp @x[9],@x[11],[$inp,#32]
248 add @x[12],@x[12],@x[13],lsl#32
249 add @x[14],@x[14],@x[15],lsl#32
250 ldp @x[13],@x[15],[$inp,#48]
262 eor @x[0],@x[0],@x[1]
263 eor @x[2],@x[2],@x[3]
264 eor @x[4],@x[4],@x[5]
265 eor @x[6],@x[6],@x[7]
266 eor @x[8],@x[8],@x[9]
267 eor @x[10],@x[10],@x[11]
268 eor @x[12],@x[12],@x[13]
269 eor @x[14],@x[14],@x[15]
271 stp @x[0],@x[2],[$out,#0] // store output
272 add @d[6],@d[6],#1 // increment counter
273 stp @x[4],@x[6],[$out,#16]
274 stp @x[8],@x[10],[$out,#32]
275 stp @x[12],@x[14],[$out,#48]
280 ldp x19,x20,[x29,#16]
282 ldp x21,x22,[x29,#32]
283 ldp x23,x24,[x29,#48]
284 ldp x25,x26,[x29,#64]
285 ldp x27,x28,[x29,#80]
287 .inst 0xd50323bf // autiasp
301 add @x[0],@x[0],@x[1],lsl#32 // pack
302 add @x[2],@x[2],@x[3],lsl#32
303 add @x[4],@x[4],@x[5],lsl#32
304 add @x[6],@x[6],@x[7],lsl#32
305 add @x[8],@x[8],@x[9],lsl#32
306 add @x[10],@x[10],@x[11],lsl#32
307 add @x[12],@x[12],@x[13],lsl#32
308 add @x[14],@x[14],@x[15],lsl#32
319 stp @x[0],@x[2],[sp,#0]
320 stp @x[4],@x[6],[sp,#16]
321 stp @x[8],@x[10],[sp,#32]
322 stp @x[12],@x[14],[sp,#48]
337 ldp x19,x20,[x29,#16]
339 ldp x21,x22,[x29,#32]
340 ldp x23,x24,[x29,#48]
341 ldp x25,x26,[x29,#64]
342 ldp x27,x28,[x29,#80]
344 .inst 0xd50323bf // autiasp
346 .size ChaCha20_ctr32,.-ChaCha20_ctr32
350 my @K = map("v$_.4s",(0..3));
351 my ($xt0,$xt1,$xt2,$xt3, $CTR,$ROT24) = map("v$_.4s",(4..9));
352 my @X = map("v$_.4s",(16,20,24,28, 17,21,25,29, 18,22,26,30, 19,23,27,31));
353 my ($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
354 $xc0,$xc1,$xc2,$xc3, $xd0,$xd1,$xd2,$xd3) = @X;
356 sub NEON_lane_ROUND {
357 my ($a0,$b0,$c0,$d0)=@_;
358 my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
359 my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
360 my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
361 my @x=map("'$_'",@X);
364 "&add (@x[$a0],@x[$a0],@x[$b0])", # Q1
365 "&add (@x[$a1],@x[$a1],@x[$b1])", # Q2
366 "&add (@x[$a2],@x[$a2],@x[$b2])", # Q3
367 "&add (@x[$a3],@x[$a3],@x[$b3])", # Q4
368 "&eor (@x[$d0],@x[$d0],@x[$a0])",
369 "&eor (@x[$d1],@x[$d1],@x[$a1])",
370 "&eor (@x[$d2],@x[$d2],@x[$a2])",
371 "&eor (@x[$d3],@x[$d3],@x[$a3])",
372 "&rev32_16 (@x[$d0],@x[$d0])",
373 "&rev32_16 (@x[$d1],@x[$d1])",
374 "&rev32_16 (@x[$d2],@x[$d2])",
375 "&rev32_16 (@x[$d3],@x[$d3])",
377 "&add (@x[$c0],@x[$c0],@x[$d0])",
378 "&add (@x[$c1],@x[$c1],@x[$d1])",
379 "&add (@x[$c2],@x[$c2],@x[$d2])",
380 "&add (@x[$c3],@x[$c3],@x[$d3])",
381 "&eor ('$xt0',@x[$b0],@x[$c0])",
382 "&eor ('$xt1',@x[$b1],@x[$c1])",
383 "&eor ('$xt2',@x[$b2],@x[$c2])",
384 "&eor ('$xt3',@x[$b3],@x[$c3])",
385 "&ushr (@x[$b0],'$xt0',20)",
386 "&ushr (@x[$b1],'$xt1',20)",
387 "&ushr (@x[$b2],'$xt2',20)",
388 "&ushr (@x[$b3],'$xt3',20)",
389 "&sli (@x[$b0],'$xt0',12)",
390 "&sli (@x[$b1],'$xt1',12)",
391 "&sli (@x[$b2],'$xt2',12)",
392 "&sli (@x[$b3],'$xt3',12)",
394 "&add (@x[$a0],@x[$a0],@x[$b0])",
395 "&add (@x[$a1],@x[$a1],@x[$b1])",
396 "&add (@x[$a2],@x[$a2],@x[$b2])",
397 "&add (@x[$a3],@x[$a3],@x[$b3])",
398 "&eor ('$xt0',@x[$d0],@x[$a0])",
399 "&eor ('$xt1',@x[$d1],@x[$a1])",
400 "&eor ('$xt2',@x[$d2],@x[$a2])",
401 "&eor ('$xt3',@x[$d3],@x[$a3])",
402 "&tbl (@x[$d0],'{$xt0}','$ROT24')",
403 "&tbl (@x[$d1],'{$xt1}','$ROT24')",
404 "&tbl (@x[$d2],'{$xt2}','$ROT24')",
405 "&tbl (@x[$d3],'{$xt3}','$ROT24')",
407 "&add (@x[$c0],@x[$c0],@x[$d0])",
408 "&add (@x[$c1],@x[$c1],@x[$d1])",
409 "&add (@x[$c2],@x[$c2],@x[$d2])",
410 "&add (@x[$c3],@x[$c3],@x[$d3])",
411 "&eor ('$xt0',@x[$b0],@x[$c0])",
412 "&eor ('$xt1',@x[$b1],@x[$c1])",
413 "&eor ('$xt2',@x[$b2],@x[$c2])",
414 "&eor ('$xt3',@x[$b3],@x[$c3])",
415 "&ushr (@x[$b0],'$xt0',25)",
416 "&ushr (@x[$b1],'$xt1',25)",
417 "&ushr (@x[$b2],'$xt2',25)",
418 "&ushr (@x[$b3],'$xt3',25)",
419 "&sli (@x[$b0],'$xt0',7)",
420 "&sli (@x[$b1],'$xt1',7)",
421 "&sli (@x[$b2],'$xt2',7)",
422 "&sli (@x[$b3],'$xt3',7)"
431 .type ChaCha20_neon,%function
435 .inst 0xd503233f // paciasp
436 stp x29,x30,[sp,#-96]!
446 b.hs .L512_or_more_neon
450 ldp @d[0],@d[1],[@x[0]] // load sigma
451 ld1 {@K[0]},[@x[0]],#16
452 ldp @d[2],@d[3],[$key] // load key
453 ldp @d[4],@d[5],[$key,#16]
454 ld1 {@K[1],@K[2]},[$key]
455 ldp @d[6],@d[7],[$ctr] // load counter
457 stp d8,d9,[sp] // meet ABI requirements
458 ld1 {$CTR,$ROT24},[@x[0]]
470 dup $xa0,@{K[0]}[0] // unpack key block
509 my @plus_one=&ROUND(0,4,8,12);
510 foreach (&NEON_lane_ROUND(0,4,8,12)) { eval; eval(shift(@plus_one)); }
512 @plus_one=&ROUND(0,5,10,15);
513 foreach (&NEON_lane_ROUND(0,5,10,15)) { eval; eval(shift(@plus_one)); }
519 zip1 $xt0,$xa0,$xa1 // transpose data
523 zip1.64 $xa0,$xt0,$xt1
524 zip2.64 $xa1,$xt0,$xt1
525 zip1.64 $xa2,$xt2,$xt3
526 zip2.64 $xa3,$xt2,$xt3
532 zip1.64 $xb0,$xt0,$xt1
533 zip2.64 $xb1,$xt0,$xt1
534 zip1.64 $xb2,$xt2,$xt3
535 zip2.64 $xb3,$xt2,$xt3
538 add.32 @x[0],@x[0],@d[0] // accumulate key block
540 add @x[1],@x[1],@d[0],lsr#32
542 add.32 @x[2],@x[2],@d[1]
544 add @x[3],@x[3],@d[1],lsr#32
545 zip1.64 $xc0,$xt0,$xt1
546 add.32 @x[4],@x[4],@d[2]
547 zip2.64 $xc1,$xt0,$xt1
548 add @x[5],@x[5],@d[2],lsr#32
549 zip1.64 $xc2,$xt2,$xt3
550 add.32 @x[6],@x[6],@d[3]
551 zip2.64 $xc3,$xt2,$xt3
552 add @x[7],@x[7],@d[3],lsr#32
555 add.32 @x[8],@x[8],@d[4]
557 add @x[9],@x[9],@d[4],lsr#32
559 add.32 @x[10],@x[10],@d[5]
561 add @x[11],@x[11],@d[5],lsr#32
562 zip1.64 $xd0,$xt0,$xt1
563 add.32 @x[12],@x[12],@d[6]
564 zip2.64 $xd1,$xt0,$xt1
565 add @x[13],@x[13],@d[6],lsr#32
566 zip1.64 $xd2,$xt2,$xt3
567 add.32 @x[14],@x[14],@d[7]
568 zip2.64 $xd3,$xt2,$xt3
569 add @x[15],@x[15],@d[7],lsr#32
573 add @x[0],@x[0],@x[1],lsl#32 // pack
574 add @x[2],@x[2],@x[3],lsl#32
575 ldp @x[1],@x[3],[$inp,#0] // load input
576 add $xa0,$xa0,@K[0] // accumulate key block
577 add @x[4],@x[4],@x[5],lsl#32
578 add @x[6],@x[6],@x[7],lsl#32
579 ldp @x[5],@x[7],[$inp,#16]
581 add @x[8],@x[8],@x[9],lsl#32
582 add @x[10],@x[10],@x[11],lsl#32
583 ldp @x[9],@x[11],[$inp,#32]
585 add @x[12],@x[12],@x[13],lsl#32
586 add @x[14],@x[14],@x[15],lsl#32
587 ldp @x[13],@x[15],[$inp,#48]
600 ld1.8 {$xt0-$xt3},[$inp],#64
601 eor @x[0],@x[0],@x[1]
603 eor @x[2],@x[2],@x[3]
605 eor @x[4],@x[4],@x[5]
607 eor @x[6],@x[6],@x[7]
609 eor @x[8],@x[8],@x[9]
612 eor @x[10],@x[10],@x[11]
614 eor @x[12],@x[12],@x[13]
616 eor @x[14],@x[14],@x[15]
618 add $CTR,$CTR,$xt0 // += 5
619 ld1.8 {$xt0-$xt3},[$inp],#64
621 stp @x[0],@x[2],[$out,#0] // store output
622 add @d[6],@d[6],#5 // increment counter
623 stp @x[4],@x[6],[$out,#16]
624 stp @x[8],@x[10],[$out,#32]
625 stp @x[12],@x[14],[$out,#48]
628 st1.8 {$xa0-$xd0},[$out],#64
633 ld1.8 {$xa0-$xd0},[$inp],#64
639 st1.8 {$xa1-$xd1},[$out],#64
644 ld1.8 {$xa1-$xd1},[$inp],#64
650 st1.8 {$xa2-$xd2},[$out],#64
656 st1.8 {$xa3-$xd3},[$out],#64
658 b.hi .Loop_outer_neon
660 ldp d8,d9,[sp] // meet ABI requirements
662 ldp x19,x20,[x29,#16]
664 ldp x21,x22,[x29,#32]
665 ldp x23,x24,[x29,#48]
666 ldp x25,x26,[x29,#64]
667 ldp x27,x28,[x29,#80]
669 .inst 0xd50323bf // autiasp
675 ldp d8,d9,[sp] // meet ABI requirements
679 add @x[0],@x[0],@x[1],lsl#32 // pack
680 add @x[2],@x[2],@x[3],lsl#32
681 ldp @x[1],@x[3],[$inp,#0] // load input
682 add @x[4],@x[4],@x[5],lsl#32
683 add @x[6],@x[6],@x[7],lsl#32
684 ldp @x[5],@x[7],[$inp,#16]
685 add @x[8],@x[8],@x[9],lsl#32
686 add @x[10],@x[10],@x[11],lsl#32
687 ldp @x[9],@x[11],[$inp,#32]
688 add @x[12],@x[12],@x[13],lsl#32
689 add @x[14],@x[14],@x[15],lsl#32
690 ldp @x[13],@x[15],[$inp,#48]
702 eor @x[0],@x[0],@x[1]
703 eor @x[2],@x[2],@x[3]
704 eor @x[4],@x[4],@x[5]
705 eor @x[6],@x[6],@x[7]
706 eor @x[8],@x[8],@x[9]
707 eor @x[10],@x[10],@x[11]
708 eor @x[12],@x[12],@x[13]
709 eor @x[14],@x[14],@x[15]
711 stp @x[0],@x[2],[$out,#0] // store output
712 add $xa0,$xa0,@K[0] // accumulate key block
713 stp @x[4],@x[6],[$out,#16]
715 stp @x[8],@x[10],[$out,#32]
717 stp @x[12],@x[14],[$out,#48]
725 ld1.8 {$xt0-$xt3},[$inp],#64
730 st1.8 {$xa0-$xd0},[$out],#64
741 ld1.8 {$xt0-$xt3},[$inp],#64
746 st1.8 {$xa1-$xd1},[$out],#64
757 ld1.8 {$xt0-$xt3},[$inp],#64
762 st1.8 {$xa2-$xd2},[$out],#64
772 st1.8 {$xa0-$xd0},[sp]
786 cbnz $len,.Loop_tail_neon
794 ldp x19,x20,[x29,#16]
796 ldp x21,x22,[x29,#32]
797 ldp x23,x24,[x29,#48]
798 ldp x25,x26,[x29,#64]
799 ldp x27,x28,[x29,#80]
801 .inst 0xd50323bf // autiasp
803 .size ChaCha20_neon,.-ChaCha20_neon
806 my @K = map("v$_.4s",(0..6));
807 my ($T0,$T1,$T2,$T3,$T4,$T5)=@K;
808 my ($A0,$B0,$C0,$D0,$A1,$B1,$C1,$D1,$A2,$B2,$C2,$D2,
809 $A3,$B3,$C3,$D3,$A4,$B4,$C4,$D4,$A5,$B5,$C5,$D5) = map("v$_.4s",(8..31));
815 my ($a,$b,$c,$d,$t)=@_;
818 "&add ('$a','$a','$b')",
819 "&eor ('$d','$d','$a')",
820 "&rev32_16 ('$d','$d')", # vrot ($d,16)
822 "&add ('$c','$c','$d')",
823 "&eor ('$t','$b','$c')",
824 "&ushr ('$b','$t',20)",
825 "&sli ('$b','$t',12)",
827 "&add ('$a','$a','$b')",
828 "&eor ('$d','$d','$a')",
829 "&tbl ('$d','{$d}','$rot24')",
831 "&add ('$c','$c','$d')",
832 "&eor ('$t','$b','$c')",
833 "&ushr ('$b','$t',25)",
834 "&sli ('$b','$t',7)",
836 "&ext ('$c','$c','$c',8)",
837 "&ext ('$d','$d','$d',$odd?4:12)",
838 "&ext ('$b','$b','$b',$odd?12:4)"
843 .type ChaCha20_512_neon,%function
846 .inst 0xd503233f // paciasp
847 stp x29,x30,[sp,#-96]!
861 ldp @d[0],@d[1],[@x[0]] // load sigma
862 ld1 {@K[0]},[@x[0]],#16
863 ldp @d[2],@d[3],[$key] // load key
864 ldp @d[4],@d[5],[$key,#16]
865 ld1 {@K[1],@K[2]},[$key]
866 ldp @d[6],@d[7],[$ctr] // load counter
868 ld1 {$ONE}[0],[@x[0]]
869 add $key,@x[0],#16 // .Lrot24
879 add @K[3],@K[3],$ONE // += 1
880 stp @K[0],@K[1],[sp,#0] // off-load key block, invariant part
881 add @K[3],@K[3],$ONE // not typo
886 shl $ONE,$ONE,#2 // 1 -> 4
888 stp d8,d9,[sp,#128+0] // meet ABI requirements
889 stp d10,d11,[sp,#128+16]
890 stp d12,d13,[sp,#128+32]
891 stp d14,d15,[sp,#128+48]
893 sub $len,$len,#512 // not typo
895 .Loop_outer_512_neon:
903 mov.32 @x[0],@d[0] // unpack key block
926 add $D4,$D0,$ONE // +4
928 add $D5,$D1,$ONE // +4
935 stp @K[3],@K[4],[sp,#48] // off-load key block, variable part
937 stp @K[5],@K[6],[sp,#80]
945 my @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0);
946 my @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0);
947 my @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0);
948 my @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,0);
949 my @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,0);
950 my @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,0);
951 my @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15));
952 my $diff = ($#thread0+1)*6 - $#thread67 - 1;
956 eval; eval(shift(@thread67));
957 eval(shift(@thread1)); eval(shift(@thread67));
958 eval(shift(@thread2)); eval(shift(@thread67));
959 eval(shift(@thread3)); eval(shift(@thread67));
960 eval(shift(@thread4)); eval(shift(@thread67));
961 eval(shift(@thread5)); eval(shift(@thread67));
964 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1);
965 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1);
966 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1);
967 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,1);
968 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,1);
969 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,1);
970 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15));
973 eval; eval(shift(@thread67));
974 eval(shift(@thread1)); eval(shift(@thread67));
975 eval(shift(@thread2)); eval(shift(@thread67));
976 eval(shift(@thread3)); eval(shift(@thread67));
977 eval(shift(@thread4)); eval(shift(@thread67));
978 eval(shift(@thread5)); eval(shift(@thread67));
981 cbnz $ctr,.Loop_upper_neon
983 add.32 @x[0],@x[0],@d[0] // accumulate key block
984 add @x[1],@x[1],@d[0],lsr#32
985 add.32 @x[2],@x[2],@d[1]
986 add @x[3],@x[3],@d[1],lsr#32
987 add.32 @x[4],@x[4],@d[2]
988 add @x[5],@x[5],@d[2],lsr#32
989 add.32 @x[6],@x[6],@d[3]
990 add @x[7],@x[7],@d[3],lsr#32
991 add.32 @x[8],@x[8],@d[4]
992 add @x[9],@x[9],@d[4],lsr#32
993 add.32 @x[10],@x[10],@d[5]
994 add @x[11],@x[11],@d[5],lsr#32
995 add.32 @x[12],@x[12],@d[6]
996 add @x[13],@x[13],@d[6],lsr#32
997 add.32 @x[14],@x[14],@d[7]
998 add @x[15],@x[15],@d[7],lsr#32
1000 add @x[0],@x[0],@x[1],lsl#32 // pack
1001 add @x[2],@x[2],@x[3],lsl#32
1002 ldp @x[1],@x[3],[$inp,#0] // load input
1003 add @x[4],@x[4],@x[5],lsl#32
1004 add @x[6],@x[6],@x[7],lsl#32
1005 ldp @x[5],@x[7],[$inp,#16]
1006 add @x[8],@x[8],@x[9],lsl#32
1007 add @x[10],@x[10],@x[11],lsl#32
1008 ldp @x[9],@x[11],[$inp,#32]
1009 add @x[12],@x[12],@x[13],lsl#32
1010 add @x[14],@x[14],@x[15],lsl#32
1011 ldp @x[13],@x[15],[$inp,#48]
1013 #ifdef __AARCH64EB__
1023 eor @x[0],@x[0],@x[1]
1024 eor @x[2],@x[2],@x[3]
1025 eor @x[4],@x[4],@x[5]
1026 eor @x[6],@x[6],@x[7]
1027 eor @x[8],@x[8],@x[9]
1028 eor @x[10],@x[10],@x[11]
1029 eor @x[12],@x[12],@x[13]
1030 eor @x[14],@x[14],@x[15]
1032 stp @x[0],@x[2],[$out,#0] // store output
1033 add @d[6],@d[6],#1 // increment counter
1034 mov.32 @x[0],@d[0] // unpack key block
1036 stp @x[4],@x[6],[$out,#16]
1039 stp @x[8],@x[10],[$out,#32]
1042 stp @x[12],@x[14],[$out,#48]
1049 lsr @x[11],@d[5],#32
1051 lsr @x[13],@d[6],#32
1053 lsr @x[15],@d[7],#32
1059 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0);
1060 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0);
1061 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0);
1062 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,0);
1063 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,0);
1064 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,0);
1065 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15));
1067 foreach (@thread0) {
1068 eval; eval(shift(@thread67));
1069 eval(shift(@thread1)); eval(shift(@thread67));
1070 eval(shift(@thread2)); eval(shift(@thread67));
1071 eval(shift(@thread3)); eval(shift(@thread67));
1072 eval(shift(@thread4)); eval(shift(@thread67));
1073 eval(shift(@thread5)); eval(shift(@thread67));
1076 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1);
1077 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1);
1078 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1);
1079 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,1);
1080 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,1);
1081 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,1);
1082 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15));
1084 foreach (@thread0) {
1085 eval; eval(shift(@thread67));
1086 eval(shift(@thread1)); eval(shift(@thread67));
1087 eval(shift(@thread2)); eval(shift(@thread67));
1088 eval(shift(@thread3)); eval(shift(@thread67));
1089 eval(shift(@thread4)); eval(shift(@thread67));
1090 eval(shift(@thread5)); eval(shift(@thread67));
1093 cbnz $ctr,.Loop_lower_neon
1095 add.32 @x[0],@x[0],@d[0] // accumulate key block
1096 ldp @K[0],@K[1],[sp,#0]
1097 add @x[1],@x[1],@d[0],lsr#32
1098 ldp @K[2],@K[3],[sp,#32]
1099 add.32 @x[2],@x[2],@d[1]
1100 ldp @K[4],@K[5],[sp,#64]
1101 add @x[3],@x[3],@d[1],lsr#32
1104 add.32 @x[4],@x[4],@d[2]
1106 add @x[5],@x[5],@d[2],lsr#32
1108 add.32 @x[6],@x[6],@d[3]
1110 add @x[7],@x[7],@d[3],lsr#32
1112 add.32 @x[8],@x[8],@d[4]
1114 add @x[9],@x[9],@d[4],lsr#32
1116 add.32 @x[10],@x[10],@d[5]
1118 add @x[11],@x[11],@d[5],lsr#32
1120 add.32 @x[12],@x[12],@d[6]
1122 add @x[13],@x[13],@d[6],lsr#32
1124 add.32 @x[14],@x[14],@d[7]
1126 add @x[15],@x[15],@d[7],lsr#32
1127 add $D4,$D4,$ONE // +4
1128 add @x[0],@x[0],@x[1],lsl#32 // pack
1129 add $D5,$D5,$ONE // +4
1130 add @x[2],@x[2],@x[3],lsl#32
1132 ldp @x[1],@x[3],[$inp,#0] // load input
1134 add @x[4],@x[4],@x[5],lsl#32
1136 add @x[6],@x[6],@x[7],lsl#32
1138 ldp @x[5],@x[7],[$inp,#16]
1140 add @x[8],@x[8],@x[9],lsl#32
1142 add @x[10],@x[10],@x[11],lsl#32
1144 ldp @x[9],@x[11],[$inp,#32]
1146 add @x[12],@x[12],@x[13],lsl#32
1148 add @x[14],@x[14],@x[15],lsl#32
1150 ldp @x[13],@x[15],[$inp,#48]
1155 #ifdef __AARCH64EB__
1165 ld1.8 {$T0-$T3},[$inp],#64
1166 eor @x[0],@x[0],@x[1]
1167 eor @x[2],@x[2],@x[3]
1168 eor @x[4],@x[4],@x[5]
1169 eor @x[6],@x[6],@x[7]
1170 eor @x[8],@x[8],@x[9]
1172 eor @x[10],@x[10],@x[11]
1174 eor @x[12],@x[12],@x[13]
1176 eor @x[14],@x[14],@x[15]
1178 ld1.8 {$T0-$T3},[$inp],#64
1180 stp @x[0],@x[2],[$out,#0] // store output
1181 add @d[6],@d[6],#7 // increment counter
1182 stp @x[4],@x[6],[$out,#16]
1183 stp @x[8],@x[10],[$out,#32]
1184 stp @x[12],@x[14],[$out,#48]
1186 st1.8 {$A0-$D0},[$out],#64
1188 ld1.8 {$A0-$D0},[$inp],#64
1193 st1.8 {$A1-$D1},[$out],#64
1195 ld1.8 {$A1-$D1},[$inp],#64
1197 ldp @K[0],@K[1],[sp,#0]
1199 ldp @K[2],@K[3],[sp,#32]
1202 st1.8 {$A2-$D2},[$out],#64
1204 ld1.8 {$A2-$D2},[$inp],#64
1209 st1.8 {$A3-$D3},[$out],#64
1211 ld1.8 {$A3-$D3},[$inp],#64
1216 st1.8 {$A4-$D4},[$out],#64
1218 shl $A0,$ONE,#1 // 4 -> 8
1223 st1.8 {$A5-$D5},[$out],#64
1225 add @K[3],@K[3],$A0 // += 8
1230 b.hs .Loop_outer_512_neon
1233 ushr $ONE,$ONE,#1 // 4 -> 2
1235 ldp d10,d11,[sp,#128+16] // meet ABI requirements
1236 ldp d12,d13,[sp,#128+32]
1237 ldp d14,d15,[sp,#128+48]
1239 stp @K[0],@K[0],[sp,#0] // wipe off-load area
1240 stp @K[0],@K[0],[sp,#32]
1241 stp @K[0],@K[0],[sp,#64]
1243 b.eq .Ldone_512_neon
1245 sub $key,$key,#16 // .Lone
1248 sub @K[3],@K[3],$ONE // -= 2
1249 ld1 {$CTR,$ROT24},[$key]
1250 b.hs .Loop_outer_neon
1252 ldp d8,d9,[sp,#0] // meet ABI requirements
1253 eor @K[1],@K[1],@K[1]
1254 eor @K[2],@K[2],@K[2]
1255 eor @K[3],@K[3],@K[3]
1256 eor @K[4],@K[4],@K[4]
1257 eor @K[5],@K[5],@K[5]
1258 eor @K[6],@K[6],@K[6]
1262 ldp d8,d9,[sp,#128+0] // meet ABI requirements
1263 ldp x19,x20,[x29,#16]
1265 ldp x21,x22,[x29,#32]
1266 ldp x23,x24,[x29,#48]
1267 ldp x25,x26,[x29,#64]
1268 ldp x27,x28,[x29,#80]
1269 ldp x29,x30,[sp],#96
1270 .inst 0xd50323bf // autiasp
1272 .size ChaCha20_512_neon,.-ChaCha20_512_neon
1277 foreach (split("\n",$code)) {
1278 s/\`([^\`]*)\`/eval $1/geo;
1280 (s/\b([a-z]+)\.32\b/$1/ and (s/x([0-9]+)/w$1/g or 1)) or
1281 (m/\b(eor|ext|mov|tbl)\b/ and (s/\.4s/\.16b/g or 1)) or
1282 (s/\b((?:ld|st)1)\.8\b/$1/ and (s/\.4s/\.16b/g or 1)) or
1283 (m/\b(ld|st)[rp]\b/ and (s/v([0-9]+)\.4s/q$1/g or 1)) or
1284 (m/\b(dup|ld1)\b/ and (s/\.4(s}?\[[0-3]\])/.$1/g or 1)) or
1285 (s/\b(zip[12])\.64\b/$1/ and (s/\.4s/\.2d/g or 1)) or
1286 (s/\brev32\.16\b/rev32/ and (s/\.4s/\.8h/g or 1));
1288 #s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo;
1292 close STDOUT or die "error closing STDOUT: $!"; # flush