Debian Codegraph @

external/boringssl/crypto/chacha/asm/chacha-x86_64.pl

raw
1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # Written by Andy Polyakov <[email protected]> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
9 #
10 # November 2014
11 #
12 # ChaCha20 for x86_64.
13 #
14 # Performance in cycles per byte out of large buffer.
15 #
16 # IALU/gcc 4.8(i) 1xSSSE3/SSE2 4xSSSE3 8xAVX2
17 #
18 # P4 9.48/+99% -/22.7(ii) -
19 # Core2 7.83/+55% 7.90/8.08 4.35
20 # Westmere 7.19/+50% 5.60/6.70 3.00
21 # Sandy Bridge 8.31/+42% 5.45/6.76 2.72
22 # Ivy Bridge 6.71/+46% 5.40/6.49 2.41
23 # Haswell 5.92/+43% 5.20/6.45 2.42 1.23
24 # Silvermont 12.0/+33% 7.75/7.40 7.03(iii)
25 # Sledgehammer 7.28/+52% -/14.2(ii) -
26 # Bulldozer 9.66/+28% 9.85/11.1 3.06(iv)
27 # VIA Nano 10.5/+46% 6.72/8.60 6.05
28 #
29 # (i) compared to older gcc 3.x one can observe >2x improvement on
30 # most platforms;
31 # (ii) as it can be seen, SSE2 performance is too low on legacy
32 # processors; NxSSE2 results are naturally better, but not
33 # impressively better than IALU ones, which is why you won't
34 # find SSE2 code below;
35 # (iii) this is not optimal result for Atom because of MSROM
36 # limitations, SSE2 can do better, but gain is considered too
37 # low to justify the [maintenance] effort;
38 # (iv) Bulldozer actually executes 4xXOP code path that delivers 2.20;
39 #
40 # Modified from upstream OpenSSL to remove the XOP code.
41
42 $flavour = shift;
43 $output = shift;
44 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
45
46 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
47
48 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
49 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
50 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
51 die "can't locate x86_64-xlate.pl";
52
53 $avx = 2;
54
55 open OUT,"| \"$^X\" $xlate $flavour $output";
56 *STDOUT=*OUT;
57
58 # input parameter block
59 ($out,$inp,$len,$key,$counter)=("%rdi","%rsi","%rdx","%rcx","%r8");
60
61 $code.=<<___;
62 .text
63
64 .extern OPENSSL_ia32cap_P
65
66 .align 64
67 .Lzero:
68 .long 0,0,0,0
69 .Lone:
70 .long 1,0,0,0
71 .Linc:
72 .long 0,1,2,3
73 .Lfour:
74 .long 4,4,4,4
75 .Lincy:
76 .long 0,2,4,6,1,3,5,7
77 .Leight:
78 .long 8,8,8,8,8,8,8,8
79 .Lrot16:
80 .byte 0x2,0x3,0x0,0x1, 0x6,0x7,0x4,0x5, 0xa,0xb,0x8,0x9, 0xe,0xf,0xc,0xd
81 .Lrot24:
82 .byte 0x3,0x0,0x1,0x2, 0x7,0x4,0x5,0x6, 0xb,0x8,0x9,0xa, 0xf,0xc,0xd,0xe
83 .Lsigma:
84 .asciz "expand 32-byte k"
85 .asciz "ChaCha20 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
86 ___
87
88 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
89 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
90 my $arg = pop;
91 $arg = "\$$arg" if ($arg*1 eq $arg);
92 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
93 }
94
95 @x=("%eax","%ebx","%ecx","%edx",map("%r${_}d",(8..11)),
96 "%nox","%nox","%nox","%nox",map("%r${_}d",(12..15)));
97 @t=("%esi","%edi");
98
99 sub ROUND { # critical path is 24 cycles per round
100 my ($a0,$b0,$c0,$d0)=@_;
101 my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
102 my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
103 my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
104 my ($xc,$xc_)=map("\"$_\"",@t);
105 my @x=map("\"$_\"",@x);
106
107 # Consider order in which variables are addressed by their
108 # index:
109 #
110 # a b c d
111 #
112 # 0 4 8 12 < even round
113 # 1 5 9 13
114 # 2 6 10 14
115 # 3 7 11 15
116 # 0 5 10 15 < odd round
117 # 1 6 11 12
118 # 2 7 8 13
119 # 3 4 9 14
120 #
121 # 'a', 'b' and 'd's are permanently allocated in registers,
122 # @x[0..7,12..15], while 'c's are maintained in memory. If
123 # you observe 'c' column, you'll notice that pair of 'c's is
124 # invariant between rounds. This means that we have to reload
125 # them once per round, in the middle. This is why you'll see
126 # bunch of 'c' stores and loads in the middle, but none in
127 # the beginning or end.
128
129 # Normally instructions would be interleaved to favour in-order
130 # execution. Generally out-of-order cores manage it gracefully,
131 # but not this time for some reason. As in-order execution
132 # cores are dying breed, old Atom is the only one around,
133 # instructions are left uninterleaved. Besides, Atom is better
134 # off executing 1xSSSE3 code anyway...
135
136 (
137 "&add (@x[$a0],@x[$b0])", # Q1
138 "&xor (@x[$d0],@x[$a0])",
139 "&rol (@x[$d0],16)",
140 "&add (@x[$a1],@x[$b1])", # Q2
141 "&xor (@x[$d1],@x[$a1])",
142 "&rol (@x[$d1],16)",
143
144 "&add ($xc,@x[$d0])",
145 "&xor (@x[$b0],$xc)",
146 "&rol (@x[$b0],12)",
147 "&add ($xc_,@x[$d1])",
148 "&xor (@x[$b1],$xc_)",
149 "&rol (@x[$b1],12)",
150
151 "&add (@x[$a0],@x[$b0])",
152 "&xor (@x[$d0],@x[$a0])",
153 "&rol (@x[$d0],8)",
154 "&add (@x[$a1],@x[$b1])",
155 "&xor (@x[$d1],@x[$a1])",
156 "&rol (@x[$d1],8)",
157
158 "&add ($xc,@x[$d0])",
159 "&xor (@x[$b0],$xc)",
160 "&rol (@x[$b0],7)",
161 "&add ($xc_,@x[$d1])",
162 "&xor (@x[$b1],$xc_)",
163 "&rol (@x[$b1],7)",
164
165 "&mov (\"4*$c0(%rsp)\",$xc)", # reload pair of 'c's
166 "&mov (\"4*$c1(%rsp)\",$xc_)",
167 "&mov ($xc,\"4*$c2(%rsp)\")",
168 "&mov ($xc_,\"4*$c3(%rsp)\")",
169
170 "&add (@x[$a2],@x[$b2])", # Q3
171 "&xor (@x[$d2],@x[$a2])",
172 "&rol (@x[$d2],16)",
173 "&add (@x[$a3],@x[$b3])", # Q4
174 "&xor (@x[$d3],@x[$a3])",
175 "&rol (@x[$d3],16)",
176
177 "&add ($xc,@x[$d2])",
178 "&xor (@x[$b2],$xc)",
179 "&rol (@x[$b2],12)",
180 "&add ($xc_,@x[$d3])",
181 "&xor (@x[$b3],$xc_)",
182 "&rol (@x[$b3],12)",
183
184 "&add (@x[$a2],@x[$b2])",
185 "&xor (@x[$d2],@x[$a2])",
186 "&rol (@x[$d2],8)",
187 "&add (@x[$a3],@x[$b3])",
188 "&xor (@x[$d3],@x[$a3])",
189 "&rol (@x[$d3],8)",
190
191 "&add ($xc,@x[$d2])",
192 "&xor (@x[$b2],$xc)",
193 "&rol (@x[$b2],7)",
194 "&add ($xc_,@x[$d3])",
195 "&xor (@x[$b3],$xc_)",
196 "&rol (@x[$b3],7)"
197 );
198 }
199
200 ########################################################################
201 # Generic code path that handles all lengths on pre-SSSE3 processors.
202 $code.=<<___;
203 .globl ChaCha20_ctr32
204 .type ChaCha20_ctr32,\@function,5
205 .align 64
206 ChaCha20_ctr32:
207 cmp \$0,$len
208 je .Lno_data
209 mov OPENSSL_ia32cap_P+4(%rip),%r10
210 test \$`1<<(41-32)`,%r10d
211 jnz .LChaCha20_ssse3
212
213 push %rbx
214 push %rbp
215 push %r12
216 push %r13
217 push %r14
218 push %r15
219 sub \$64+24,%rsp
220
221 #movdqa .Lsigma(%rip),%xmm0
222 movdqu ($key),%xmm1
223 movdqu 16($key),%xmm2
224 movdqu ($counter),%xmm3
225 movdqa .Lone(%rip),%xmm4
226
227 #movdqa %xmm0,4*0(%rsp) # key[0]
228 movdqa %xmm1,4*4(%rsp) # key[1]
229 movdqa %xmm2,4*8(%rsp) # key[2]
230 movdqa %xmm3,4*12(%rsp) # key[3]
231 mov $len,%rbp # reassign $len
232 jmp .Loop_outer
233
234 .align 32
235 .Loop_outer:
236 mov \$0x61707865,@x[0] # 'expa'
237 mov \$0x3320646e,@x[1] # 'nd 3'
238 mov \$0x79622d32,@x[2] # '2-by'
239 mov \$0x6b206574,@x[3] # 'te k'
240 mov 4*4(%rsp),@x[4]
241 mov 4*5(%rsp),@x[5]
242 mov 4*6(%rsp),@x[6]
243 mov 4*7(%rsp),@x[7]
244 movd %xmm3,@x[12]
245 mov 4*13(%rsp),@x[13]
246 mov 4*14(%rsp),@x[14]
247 mov 4*15(%rsp),@x[15]
248
249 mov %rbp,64+0(%rsp) # save len
250 mov \$10,%ebp
251 mov $inp,64+8(%rsp) # save inp
252 movq %xmm2,%rsi # "@x[8]"
253 mov $out,64+16(%rsp) # save out
254 mov %rsi,%rdi
255 shr \$32,%rdi # "@x[9]"
256 jmp .Loop
257
258 .align 32
259 .Loop:
260 ___
261 foreach (&ROUND (0, 4, 8,12)) { eval; }
262 foreach (&ROUND (0, 5,10,15)) { eval; }
263 &dec ("%ebp");
264 &jnz (".Loop");
265
266 $code.=<<___;
267 mov @t[1],4*9(%rsp) # modulo-scheduled
268 mov @t[0],4*8(%rsp)
269 mov 64(%rsp),%rbp # load len
270 movdqa %xmm2,%xmm1
271 mov 64+8(%rsp),$inp # load inp
272 paddd %xmm4,%xmm3 # increment counter
273 mov 64+16(%rsp),$out # load out
274
275 add \$0x61707865,@x[0] # 'expa'
276 add \$0x3320646e,@x[1] # 'nd 3'
277 add \$0x79622d32,@x[2] # '2-by'
278 add \$0x6b206574,@x[3] # 'te k'
279 add 4*4(%rsp),@x[4]
280 add 4*5(%rsp),@x[5]
281 add 4*6(%rsp),@x[6]
282 add 4*7(%rsp),@x[7]
283 add 4*12(%rsp),@x[12]
284 add 4*13(%rsp),@x[13]
285 add 4*14(%rsp),@x[14]
286 add 4*15(%rsp),@x[15]
287 paddd 4*8(%rsp),%xmm1
288
289 cmp \$64,%rbp
290 jb .Ltail
291
292 xor 4*0($inp),@x[0] # xor with input
293 xor 4*1($inp),@x[1]
294 xor 4*2($inp),@x[2]
295 xor 4*3($inp),@x[3]
296 xor 4*4($inp),@x[4]
297 xor 4*5($inp),@x[5]
298 xor 4*6($inp),@x[6]
299 xor 4*7($inp),@x[7]
300 movdqu 4*8($inp),%xmm0
301 xor 4*12($inp),@x[12]
302 xor 4*13($inp),@x[13]
303 xor 4*14($inp),@x[14]
304 xor 4*15($inp),@x[15]
305 lea 4*16($inp),$inp # inp+=64
306 pxor %xmm1,%xmm0
307
308 movdqa %xmm2,4*8(%rsp)
309 movd %xmm3,4*12(%rsp)
310
311 mov @x[0],4*0($out) # write output
312 mov @x[1],4*1($out)
313 mov @x[2],4*2($out)
314 mov @x[3],4*3($out)
315 mov @x[4],4*4($out)
316 mov @x[5],4*5($out)
317 mov @x[6],4*6($out)
318 mov @x[7],4*7($out)
319 movdqu %xmm0,4*8($out)
320 mov @x[12],4*12($out)
321 mov @x[13],4*13($out)
322 mov @x[14],4*14($out)
323 mov @x[15],4*15($out)
324 lea 4*16($out),$out # out+=64
325
326 sub \$64,%rbp
327 jnz .Loop_outer
328
329 jmp .Ldone
330
331 .align 16
332 .Ltail:
333 mov @x[0],4*0(%rsp)
334 mov @x[1],4*1(%rsp)
335 xor %rbx,%rbx
336 mov @x[2],4*2(%rsp)
337 mov @x[3],4*3(%rsp)
338 mov @x[4],4*4(%rsp)
339 mov @x[5],4*5(%rsp)
340 mov @x[6],4*6(%rsp)
341 mov @x[7],4*7(%rsp)
342 movdqa %xmm1,4*8(%rsp)
343 mov @x[12],4*12(%rsp)
344 mov @x[13],4*13(%rsp)
345 mov @x[14],4*14(%rsp)
346 mov @x[15],4*15(%rsp)
347
348 .Loop_tail:
349 movzb ($inp,%rbx),%eax
350 movzb (%rsp,%rbx),%edx
351 lea 1(%rbx),%rbx
352 xor %edx,%eax
353 mov %al,-1($out,%rbx)
354 dec %rbp
355 jnz .Loop_tail
356
357 .Ldone:
358 add \$64+24,%rsp
359 pop %r15
360 pop %r14
361 pop %r13
362 pop %r12
363 pop %rbp
364 pop %rbx
365 .Lno_data:
366 ret
367 .size ChaCha20_ctr32,.-ChaCha20_ctr32
368 ___
369
370 ########################################################################
371 # SSSE3 code path that handles shorter lengths
372 {
373 my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("%xmm$_",(0..7));
374
375 sub SSSE3ROUND { # critical path is 20 "SIMD ticks" per round
376 &paddd ($a,$b);
377 &pxor ($d,$a);
378 &pshufb ($d,$rot16);
379
380 &paddd ($c,$d);
381 &pxor ($b,$c);
382 &movdqa ($t,$b);
383 &psrld ($b,20);
384 &pslld ($t,12);
385 &por ($b,$t);
386
387 &paddd ($a,$b);
388 &pxor ($d,$a);
389 &pshufb ($d,$rot24);
390
391 &paddd ($c,$d);
392 &pxor ($b,$c);
393 &movdqa ($t,$b);
394 &psrld ($b,25);
395 &pslld ($t,7);
396 &por ($b,$t);
397 }
398
399 my $xframe = $win64 ? 32+32+8 : 24;
400
401 $code.=<<___;
402 .type ChaCha20_ssse3,\@function,5
403 .align 32
404 ChaCha20_ssse3:
405 .LChaCha20_ssse3:
406 ___
407 $code.=<<___;
408 cmp \$128,$len # we might throw away some data,
409 ja .LChaCha20_4x # but overall it won't be slower
410
411 .Ldo_sse3_after_all:
412 push %rbx
413 push %rbp
414 push %r12
415 push %r13
416 push %r14
417 push %r15
418
419 sub \$64+$xframe,%rsp
420 ___
421 $code.=<<___ if ($win64);
422 movaps %xmm6,64+32(%rsp)
423 movaps %xmm7,64+48(%rsp)
424 ___
425 $code.=<<___;
426 movdqa .Lsigma(%rip),$a
427 movdqu ($key),$b
428 movdqu 16($key),$c
429 movdqu ($counter),$d
430 movdqa .Lrot16(%rip),$rot16
431 movdqa .Lrot24(%rip),$rot24
432
433 movdqa $a,0x00(%rsp)
434 movdqa $b,0x10(%rsp)
435 movdqa $c,0x20(%rsp)
436 movdqa $d,0x30(%rsp)
437 mov \$10,%ebp
438 jmp .Loop_ssse3
439
440 .align 32
441 .Loop_outer_ssse3:
442 movdqa .Lone(%rip),$d
443 movdqa 0x00(%rsp),$a
444 movdqa 0x10(%rsp),$b
445 movdqa 0x20(%rsp),$c
446 paddd 0x30(%rsp),$d
447 mov \$10,%ebp
448 movdqa $d,0x30(%rsp)
449 jmp .Loop_ssse3
450
451 .align 32
452 .Loop_ssse3:
453 ___
454 &SSSE3ROUND();
455 &pshufd ($c,$c,0b01001110);
456 &pshufd ($b,$b,0b00111001);
457 &pshufd ($d,$d,0b10010011);
458 &nop ();
459
460 &SSSE3ROUND();
461 &pshufd ($c,$c,0b01001110);
462 &pshufd ($b,$b,0b10010011);
463 &pshufd ($d,$d,0b00111001);
464
465 &dec ("%ebp");
466 &jnz (".Loop_ssse3");
467
468 $code.=<<___;
469 paddd 0x00(%rsp),$a
470 paddd 0x10(%rsp),$b
471 paddd 0x20(%rsp),$c
472 paddd 0x30(%rsp),$d
473
474 cmp \$64,$len
475 jb .Ltail_ssse3
476
477 movdqu 0x00($inp),$t
478 movdqu 0x10($inp),$t1
479 pxor $t,$a # xor with input
480 movdqu 0x20($inp),$t
481 pxor $t1,$b
482 movdqu 0x30($inp),$t1
483 lea 0x40($inp),$inp # inp+=64
484 pxor $t,$c
485 pxor $t1,$d
486
487 movdqu $a,0x00($out) # write output
488 movdqu $b,0x10($out)
489 movdqu $c,0x20($out)
490 movdqu $d,0x30($out)
491 lea 0x40($out),$out # out+=64
492
493 sub \$64,$len
494 jnz .Loop_outer_ssse3
495
496 jmp .Ldone_ssse3
497
498 .align 16
499 .Ltail_ssse3:
500 movdqa $a,0x00(%rsp)
501 movdqa $b,0x10(%rsp)
502 movdqa $c,0x20(%rsp)
503 movdqa $d,0x30(%rsp)
504 xor %rbx,%rbx
505
506 .Loop_tail_ssse3:
507 movzb ($inp,%rbx),%eax
508 movzb (%rsp,%rbx),%ecx
509 lea 1(%rbx),%rbx
510 xor %ecx,%eax
511 mov %al,-1($out,%rbx)
512 dec $len
513 jnz .Loop_tail_ssse3
514
515 .Ldone_ssse3:
516 ___
517 $code.=<<___ if ($win64);
518 movaps 64+32(%rsp),%xmm6
519 movaps 64+48(%rsp),%xmm7
520 ___
521 $code.=<<___;
522 add \$64+$xframe,%rsp
523 pop %r15
524 pop %r14
525 pop %r13
526 pop %r12
527 pop %rbp
528 pop %rbx
529 ret
530 .size ChaCha20_ssse3,.-ChaCha20_ssse3
531 ___
532 }
533
534 ########################################################################
535 # SSSE3 code path that handles longer messages.
536 {
537 # assign variables to favor Atom front-end
538 my ($xd0,$xd1,$xd2,$xd3, $xt0,$xt1,$xt2,$xt3,
539 $xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3)=map("%xmm$_",(0..15));
540 my @xx=($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
541 "%nox","%nox","%nox","%nox", $xd0,$xd1,$xd2,$xd3);
542
543 sub SSSE3_lane_ROUND {
544 my ($a0,$b0,$c0,$d0)=@_;
545 my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
546 my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
547 my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
548 my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
549 my @x=map("\"$_\"",@xx);
550
551 # Consider order in which variables are addressed by their
552 # index:
553 #
554 # a b c d
555 #
556 # 0 4 8 12 < even round
557 # 1 5 9 13
558 # 2 6 10 14
559 # 3 7 11 15
560 # 0 5 10 15 < odd round
561 # 1 6 11 12
562 # 2 7 8 13
563 # 3 4 9 14
564 #
565 # 'a', 'b' and 'd's are permanently allocated in registers,
566 # @x[0..7,12..15], while 'c's are maintained in memory. If
567 # you observe 'c' column, you'll notice that pair of 'c's is
568 # invariant between rounds. This means that we have to reload
569 # them once per round, in the middle. This is why you'll see
570 # bunch of 'c' stores and loads in the middle, but none in
571 # the beginning or end.
572
573 (
574 "&paddd (@x[$a0],@x[$b0])", # Q1
575 "&paddd (@x[$a1],@x[$b1])", # Q2
576 "&pxor (@x[$d0],@x[$a0])",
577 "&pxor (@x[$d1],@x[$a1])",
578 "&pshufb (@x[$d0],$t1)",
579 "&pshufb (@x[$d1],$t1)",
580
581 "&paddd ($xc,@x[$d0])",
582 "&paddd ($xc_,@x[$d1])",
583 "&pxor (@x[$b0],$xc)",
584 "&pxor (@x[$b1],$xc_)",
585 "&movdqa ($t0,@x[$b0])",
586 "&pslld (@x[$b0],12)",
587 "&psrld ($t0,20)",
588 "&movdqa ($t1,@x[$b1])",
589 "&pslld (@x[$b1],12)",
590 "&por (@x[$b0],$t0)",
591 "&psrld ($t1,20)",
592 "&movdqa ($t0,'(%r11)')", # .Lrot24(%rip)
593 "&por (@x[$b1],$t1)",
594
595 "&paddd (@x[$a0],@x[$b0])",
596 "&paddd (@x[$a1],@x[$b1])",
597 "&pxor (@x[$d0],@x[$a0])",
598 "&pxor (@x[$d1],@x[$a1])",
599 "&pshufb (@x[$d0],$t0)",
600 "&pshufb (@x[$d1],$t0)",
601
602 "&paddd ($xc,@x[$d0])",
603 "&paddd ($xc_,@x[$d1])",
604 "&pxor (@x[$b0],$xc)",
605 "&pxor (@x[$b1],$xc_)",
606 "&movdqa ($t1,@x[$b0])",
607 "&pslld (@x[$b0],7)",
608 "&psrld ($t1,25)",
609 "&movdqa ($t0,@x[$b1])",
610 "&pslld (@x[$b1],7)",
611 "&por (@x[$b0],$t1)",
612 "&psrld ($t0,25)",
613 "&movdqa ($t1,'(%r10)')", # .Lrot16(%rip)
614 "&por (@x[$b1],$t0)",
615
616 "&movdqa (\"`16*($c0-8)`(%rsp)\",$xc)", # reload pair of 'c's
617 "&movdqa (\"`16*($c1-8)`(%rsp)\",$xc_)",
618 "&movdqa ($xc,\"`16*($c2-8)`(%rsp)\")",
619 "&movdqa ($xc_,\"`16*($c3-8)`(%rsp)\")",
620
621 "&paddd (@x[$a2],@x[$b2])", # Q3
622 "&paddd (@x[$a3],@x[$b3])", # Q4
623 "&pxor (@x[$d2],@x[$a2])",
624 "&pxor (@x[$d3],@x[$a3])",
625 "&pshufb (@x[$d2],$t1)",
626 "&pshufb (@x[$d3],$t1)",
627
628 "&paddd ($xc,@x[$d2])",
629 "&paddd ($xc_,@x[$d3])",
630 "&pxor (@x[$b2],$xc)",
631 "&pxor (@x[$b3],$xc_)",
632 "&movdqa ($t0,@x[$b2])",
633 "&pslld (@x[$b2],12)",
634 "&psrld ($t0,20)",
635 "&movdqa ($t1,@x[$b3])",
636 "&pslld (@x[$b3],12)",
637 "&por (@x[$b2],$t0)",
638 "&psrld ($t1,20)",
639 "&movdqa ($t0,'(%r11)')", # .Lrot24(%rip)
640 "&por (@x[$b3],$t1)",
641
642 "&paddd (@x[$a2],@x[$b2])",
643 "&paddd (@x[$a3],@x[$b3])",
644 "&pxor (@x[$d2],@x[$a2])",
645 "&pxor (@x[$d3],@x[$a3])",
646 "&pshufb (@x[$d2],$t0)",
647 "&pshufb (@x[$d3],$t0)",
648
649 "&paddd ($xc,@x[$d2])",
650 "&paddd ($xc_,@x[$d3])",
651 "&pxor (@x[$b2],$xc)",
652 "&pxor (@x[$b3],$xc_)",
653 "&movdqa ($t1,@x[$b2])",
654 "&pslld (@x[$b2],7)",
655 "&psrld ($t1,25)",
656 "&movdqa ($t0,@x[$b3])",
657 "&pslld (@x[$b3],7)",
658 "&por (@x[$b2],$t1)",
659 "&psrld ($t0,25)",
660 "&movdqa ($t1,'(%r10)')", # .Lrot16(%rip)
661 "&por (@x[$b3],$t0)"
662 );
663 }
664
665 my $xframe = $win64 ? 0xa0 : 0;
666
667 $code.=<<___;
668 .type ChaCha20_4x,\@function,5
669 .align 32
670 ChaCha20_4x:
671 .LChaCha20_4x:
672 mov %r10,%r11
673 ___
674 $code.=<<___ if ($avx>1);
675 shr \$32,%r10 # OPENSSL_ia32cap_P+8
676 test \$`1<<5`,%r10 # test AVX2
677 jnz .LChaCha20_8x
678 ___
679 $code.=<<___;
680 cmp \$192,$len
681 ja .Lproceed4x
682
683 and \$`1<<26|1<<22`,%r11 # isolate XSAVE+MOVBE
684 cmp \$`1<<22`,%r11 # check for MOVBE without XSAVE
685 je .Ldo_sse3_after_all # to detect Atom
686
687 .Lproceed4x:
688 lea -0x78(%rsp),%r11
689 sub \$0x148+$xframe,%rsp
690 ___
691 ################ stack layout
692 # +0x00 SIMD equivalent of @x[8-12]
693 # ...
694 # +0x40 constant copy of key[0-2] smashed by lanes
695 # ...
696 # +0x100 SIMD counters (with nonce smashed by lanes)
697 # ...
698 # +0x140
699 $code.=<<___ if ($win64);
700 movaps %xmm6,-0x30(%r11)
701 movaps %xmm7,-0x20(%r11)
702 movaps %xmm8,-0x10(%r11)
703 movaps %xmm9,0x00(%r11)
704 movaps %xmm10,0x10(%r11)
705 movaps %xmm11,0x20(%r11)
706 movaps %xmm12,0x30(%r11)
707 movaps %xmm13,0x40(%r11)
708 movaps %xmm14,0x50(%r11)
709 movaps %xmm15,0x60(%r11)
710 ___
711 $code.=<<___;
712 movdqa .Lsigma(%rip),$xa3 # key[0]
713 movdqu ($key),$xb3 # key[1]
714 movdqu 16($key),$xt3 # key[2]
715 movdqu ($counter),$xd3 # key[3]
716 lea 0x100(%rsp),%rcx # size optimization
717 lea .Lrot16(%rip),%r10
718 lea .Lrot24(%rip),%r11
719
720 pshufd \$0x00,$xa3,$xa0 # smash key by lanes...
721 pshufd \$0x55,$xa3,$xa1
722 movdqa $xa0,0x40(%rsp) # ... and offload
723 pshufd \$0xaa,$xa3,$xa2
724 movdqa $xa1,0x50(%rsp)
725 pshufd \$0xff,$xa3,$xa3
726 movdqa $xa2,0x60(%rsp)
727 movdqa $xa3,0x70(%rsp)
728
729 pshufd \$0x00,$xb3,$xb0
730 pshufd \$0x55,$xb3,$xb1
731 movdqa $xb0,0x80-0x100(%rcx)
732 pshufd \$0xaa,$xb3,$xb2
733 movdqa $xb1,0x90-0x100(%rcx)
734 pshufd \$0xff,$xb3,$xb3
735 movdqa $xb2,0xa0-0x100(%rcx)
736 movdqa $xb3,0xb0-0x100(%rcx)
737
738 pshufd \$0x00,$xt3,$xt0 # "$xc0"
739 pshufd \$0x55,$xt3,$xt1 # "$xc1"
740 movdqa $xt0,0xc0-0x100(%rcx)
741 pshufd \$0xaa,$xt3,$xt2 # "$xc2"
742 movdqa $xt1,0xd0-0x100(%rcx)
743 pshufd \$0xff,$xt3,$xt3 # "$xc3"
744 movdqa $xt2,0xe0-0x100(%rcx)
745 movdqa $xt3,0xf0-0x100(%rcx)
746
747 pshufd \$0x00,$xd3,$xd0
748 pshufd \$0x55,$xd3,$xd1
749 paddd .Linc(%rip),$xd0 # don't save counters yet
750 pshufd \$0xaa,$xd3,$xd2
751 movdqa $xd1,0x110-0x100(%rcx)
752 pshufd \$0xff,$xd3,$xd3
753 movdqa $xd2,0x120-0x100(%rcx)
754 movdqa $xd3,0x130-0x100(%rcx)
755
756 jmp .Loop_enter4x
757
758 .align 32
759 .Loop_outer4x:
760 movdqa 0x40(%rsp),$xa0 # re-load smashed key
761 movdqa 0x50(%rsp),$xa1
762 movdqa 0x60(%rsp),$xa2
763 movdqa 0x70(%rsp),$xa3
764 movdqa 0x80-0x100(%rcx),$xb0
765 movdqa 0x90-0x100(%rcx),$xb1
766 movdqa 0xa0-0x100(%rcx),$xb2
767 movdqa 0xb0-0x100(%rcx),$xb3
768 movdqa 0xc0-0x100(%rcx),$xt0 # "$xc0"
769 movdqa 0xd0-0x100(%rcx),$xt1 # "$xc1"
770 movdqa 0xe0-0x100(%rcx),$xt2 # "$xc2"
771 movdqa 0xf0-0x100(%rcx),$xt3 # "$xc3"
772 movdqa 0x100-0x100(%rcx),$xd0
773 movdqa 0x110-0x100(%rcx),$xd1
774 movdqa 0x120-0x100(%rcx),$xd2
775 movdqa 0x130-0x100(%rcx),$xd3
776 paddd .Lfour(%rip),$xd0 # next SIMD counters
777
778 .Loop_enter4x:
779 movdqa $xt2,0x20(%rsp) # SIMD equivalent of "@x[10]"
780 movdqa $xt3,0x30(%rsp) # SIMD equivalent of "@x[11]"
781 movdqa (%r10),$xt3 # .Lrot16(%rip)
782 mov \$10,%eax
783 movdqa $xd0,0x100-0x100(%rcx) # save SIMD counters
784 jmp .Loop4x
785
786 .align 32
787 .Loop4x:
788 ___
789 foreach (&SSSE3_lane_ROUND(0, 4, 8,12)) { eval; }
790 foreach (&SSSE3_lane_ROUND(0, 5,10,15)) { eval; }
791 $code.=<<___;
792 dec %eax
793 jnz .Loop4x
794
795 paddd 0x40(%rsp),$xa0 # accumulate key material
796 paddd 0x50(%rsp),$xa1
797 paddd 0x60(%rsp),$xa2
798 paddd 0x70(%rsp),$xa3
799
800 movdqa $xa0,$xt2 # "de-interlace" data
801 punpckldq $xa1,$xa0
802 movdqa $xa2,$xt3
803 punpckldq $xa3,$xa2
804 punpckhdq $xa1,$xt2
805 punpckhdq $xa3,$xt3
806 movdqa $xa0,$xa1
807 punpcklqdq $xa2,$xa0 # "a0"
808 movdqa $xt2,$xa3
809 punpcklqdq $xt3,$xt2 # "a2"
810 punpckhqdq $xa2,$xa1 # "a1"
811 punpckhqdq $xt3,$xa3 # "a3"
812 ___
813 ($xa2,$xt2)=($xt2,$xa2);
814 $code.=<<___;
815 paddd 0x80-0x100(%rcx),$xb0
816 paddd 0x90-0x100(%rcx),$xb1
817 paddd 0xa0-0x100(%rcx),$xb2
818 paddd 0xb0-0x100(%rcx),$xb3
819
820 movdqa $xa0,0x00(%rsp) # offload $xaN
821 movdqa $xa1,0x10(%rsp)
822 movdqa 0x20(%rsp),$xa0 # "xc2"
823 movdqa 0x30(%rsp),$xa1 # "xc3"
824
825 movdqa $xb0,$xt2
826 punpckldq $xb1,$xb0
827 movdqa $xb2,$xt3
828 punpckldq $xb3,$xb2
829 punpckhdq $xb1,$xt2
830 punpckhdq $xb3,$xt3
831 movdqa $xb0,$xb1
832 punpcklqdq $xb2,$xb0 # "b0"
833 movdqa $xt2,$xb3
834 punpcklqdq $xt3,$xt2 # "b2"
835 punpckhqdq $xb2,$xb1 # "b1"
836 punpckhqdq $xt3,$xb3 # "b3"
837 ___
838 ($xb2,$xt2)=($xt2,$xb2);
839 my ($xc0,$xc1,$xc2,$xc3)=($xt0,$xt1,$xa0,$xa1);
840 $code.=<<___;
841 paddd 0xc0-0x100(%rcx),$xc0
842 paddd 0xd0-0x100(%rcx),$xc1
843 paddd 0xe0-0x100(%rcx),$xc2
844 paddd 0xf0-0x100(%rcx),$xc3
845
846 movdqa $xa2,0x20(%rsp) # keep offloading $xaN
847 movdqa $xa3,0x30(%rsp)
848
849 movdqa $xc0,$xt2
850 punpckldq $xc1,$xc0
851 movdqa $xc2,$xt3
852 punpckldq $xc3,$xc2
853 punpckhdq $xc1,$xt2
854 punpckhdq $xc3,$xt3
855 movdqa $xc0,$xc1
856 punpcklqdq $xc2,$xc0 # "c0"
857 movdqa $xt2,$xc3
858 punpcklqdq $xt3,$xt2 # "c2"
859 punpckhqdq $xc2,$xc1 # "c1"
860 punpckhqdq $xt3,$xc3 # "c3"
861 ___
862 ($xc2,$xt2)=($xt2,$xc2);
863 ($xt0,$xt1)=($xa2,$xa3); # use $xaN as temporary
864 $code.=<<___;
865 paddd 0x100-0x100(%rcx),$xd0
866 paddd 0x110-0x100(%rcx),$xd1
867 paddd 0x120-0x100(%rcx),$xd2
868 paddd 0x130-0x100(%rcx),$xd3
869
870 movdqa $xd0,$xt2
871 punpckldq $xd1,$xd0
872 movdqa $xd2,$xt3
873 punpckldq $xd3,$xd2
874 punpckhdq $xd1,$xt2
875 punpckhdq $xd3,$xt3
876 movdqa $xd0,$xd1
877 punpcklqdq $xd2,$xd0 # "d0"
878 movdqa $xt2,$xd3
879 punpcklqdq $xt3,$xt2 # "d2"
880 punpckhqdq $xd2,$xd1 # "d1"
881 punpckhqdq $xt3,$xd3 # "d3"
882 ___
883 ($xd2,$xt2)=($xt2,$xd2);
884 $code.=<<___;
885 cmp \$64*4,$len
886 jb .Ltail4x
887
888 movdqu 0x00($inp),$xt0 # xor with input
889 movdqu 0x10($inp),$xt1
890 movdqu 0x20($inp),$xt2
891 movdqu 0x30($inp),$xt3
892 pxor 0x00(%rsp),$xt0 # $xaN is offloaded, remember?
893 pxor $xb0,$xt1
894 pxor $xc0,$xt2
895 pxor $xd0,$xt3
896
897 movdqu $xt0,0x00($out)
898 movdqu 0x40($inp),$xt0
899 movdqu $xt1,0x10($out)
900 movdqu 0x50($inp),$xt1
901 movdqu $xt2,0x20($out)
902 movdqu 0x60($inp),$xt2
903 movdqu $xt3,0x30($out)
904 movdqu 0x70($inp),$xt3
905 lea 0x80($inp),$inp # size optimization
906 pxor 0x10(%rsp),$xt0
907 pxor $xb1,$xt1
908 pxor $xc1,$xt2
909 pxor $xd1,$xt3
910
911 movdqu $xt0,0x40($out)
912 movdqu 0x00($inp),$xt0
913 movdqu $xt1,0x50($out)
914 movdqu 0x10($inp),$xt1
915 movdqu $xt2,0x60($out)
916 movdqu 0x20($inp),$xt2
917 movdqu $xt3,0x70($out)
918 lea 0x80($out),$out # size optimization
919 movdqu 0x30($inp),$xt3
920 pxor 0x20(%rsp),$xt0
921 pxor $xb2,$xt1
922 pxor $xc2,$xt2
923 pxor $xd2,$xt3
924
925 movdqu $xt0,0x00($out)
926 movdqu 0x40($inp),$xt0
927 movdqu $xt1,0x10($out)
928 movdqu 0x50($inp),$xt1
929 movdqu $xt2,0x20($out)
930 movdqu 0x60($inp),$xt2
931 movdqu $xt3,0x30($out)
932 movdqu 0x70($inp),$xt3
933 lea 0x80($inp),$inp # inp+=64*4
934 pxor 0x30(%rsp),$xt0
935 pxor $xb3,$xt1
936 pxor $xc3,$xt2
937 pxor $xd3,$xt3
938 movdqu $xt0,0x40($out)
939 movdqu $xt1,0x50($out)
940 movdqu $xt2,0x60($out)
941 movdqu $xt3,0x70($out)
942 lea 0x80($out),$out # out+=64*4
943
944 sub \$64*4,$len
945 jnz .Loop_outer4x
946
947 jmp .Ldone4x
948
949 .Ltail4x:
950 cmp \$192,$len
951 jae .L192_or_more4x
952 cmp \$128,$len
953 jae .L128_or_more4x
954 cmp \$64,$len
955 jae .L64_or_more4x
956
957 #movdqa 0x00(%rsp),$xt0 # $xaN is offloaded, remember?
958 xor %r10,%r10
959 #movdqa $xt0,0x00(%rsp)
960 movdqa $xb0,0x10(%rsp)
961 movdqa $xc0,0x20(%rsp)
962 movdqa $xd0,0x30(%rsp)
963 jmp .Loop_tail4x
964
965 .align 32
966 .L64_or_more4x:
967 movdqu 0x00($inp),$xt0 # xor with input
968 movdqu 0x10($inp),$xt1
969 movdqu 0x20($inp),$xt2
970 movdqu 0x30($inp),$xt3
971 pxor 0x00(%rsp),$xt0 # $xaxN is offloaded, remember?
972 pxor $xb0,$xt1
973 pxor $xc0,$xt2
974 pxor $xd0,$xt3
975 movdqu $xt0,0x00($out)
976 movdqu $xt1,0x10($out)
977 movdqu $xt2,0x20($out)
978 movdqu $xt3,0x30($out)
979 je .Ldone4x
980
981 movdqa 0x10(%rsp),$xt0 # $xaN is offloaded, remember?
982 lea 0x40($inp),$inp # inp+=64*1
983 xor %r10,%r10
984 movdqa $xt0,0x00(%rsp)
985 movdqa $xb1,0x10(%rsp)
986 lea 0x40($out),$out # out+=64*1
987 movdqa $xc1,0x20(%rsp)
988 sub \$64,$len # len-=64*1
989 movdqa $xd1,0x30(%rsp)
990 jmp .Loop_tail4x
991
992 .align 32
993 .L128_or_more4x:
994 movdqu 0x00($inp),$xt0 # xor with input
995 movdqu 0x10($inp),$xt1
996 movdqu 0x20($inp),$xt2
997 movdqu 0x30($inp),$xt3
998 pxor 0x00(%rsp),$xt0 # $xaN is offloaded, remember?
999 pxor $xb0,$xt1
1000 pxor $xc0,$xt2
1001 pxor $xd0,$xt3
1002
1003 movdqu $xt0,0x00($out)
1004 movdqu 0x40($inp),$xt0
1005 movdqu $xt1,0x10($out)
1006 movdqu 0x50($inp),$xt1
1007 movdqu $xt2,0x20($out)
1008 movdqu 0x60($inp),$xt2
1009 movdqu $xt3,0x30($out)
1010 movdqu 0x70($inp),$xt3
1011 pxor 0x10(%rsp),$xt0
1012 pxor $xb1,$xt1
1013 pxor $xc1,$xt2
1014 pxor $xd1,$xt3
1015 movdqu $xt0,0x40($out)
1016 movdqu $xt1,0x50($out)
1017 movdqu $xt2,0x60($out)
1018 movdqu $xt3,0x70($out)
1019 je .Ldone4x
1020
1021 movdqa 0x20(%rsp),$xt0 # $xaN is offloaded, remember?
1022 lea 0x80($inp),$inp # inp+=64*2
1023 xor %r10,%r10
1024 movdqa $xt0,0x00(%rsp)
1025 movdqa $xb2,0x10(%rsp)
1026 lea 0x80($out),$out # out+=64*2
1027 movdqa $xc2,0x20(%rsp)
1028 sub \$128,$len # len-=64*2
1029 movdqa $xd2,0x30(%rsp)
1030 jmp .Loop_tail4x
1031
1032 .align 32
1033 .L192_or_more4x:
1034 movdqu 0x00($inp),$xt0 # xor with input
1035 movdqu 0x10($inp),$xt1
1036 movdqu 0x20($inp),$xt2
1037 movdqu 0x30($inp),$xt3
1038 pxor 0x00(%rsp),$xt0 # $xaN is offloaded, remember?
1039 pxor $xb0,$xt1
1040 pxor $xc0,$xt2
1041 pxor $xd0,$xt3
1042
1043 movdqu $xt0,0x00($out)
1044 movdqu 0x40($inp),$xt0
1045 movdqu $xt1,0x10($out)
1046 movdqu 0x50($inp),$xt1
1047 movdqu $xt2,0x20($out)
1048 movdqu 0x60($inp),$xt2
1049 movdqu $xt3,0x30($out)
1050 movdqu 0x70($inp),$xt3
1051 lea 0x80($inp),$inp # size optimization
1052 pxor 0x10(%rsp),$xt0
1053 pxor $xb1,$xt1
1054 pxor $xc1,$xt2
1055 pxor $xd1,$xt3
1056
1057 movdqu $xt0,0x40($out)
1058 movdqu 0x00($inp),$xt0
1059 movdqu $xt1,0x50($out)
1060 movdqu 0x10($inp),$xt1
1061 movdqu $xt2,0x60($out)
1062 movdqu 0x20($inp),$xt2
1063 movdqu $xt3,0x70($out)
1064 lea 0x80($out),$out # size optimization
1065 movdqu 0x30($inp),$xt3
1066 pxor 0x20(%rsp),$xt0
1067 pxor $xb2,$xt1
1068 pxor $xc2,$xt2
1069 pxor $xd2,$xt3
1070 movdqu $xt0,0x00($out)
1071 movdqu $xt1,0x10($out)
1072 movdqu $xt2,0x20($out)
1073 movdqu $xt3,0x30($out)
1074 je .Ldone4x
1075
1076 movdqa 0x30(%rsp),$xt0 # $xaN is offloaded, remember?
1077 lea 0x40($inp),$inp # inp+=64*3
1078 xor %r10,%r10
1079 movdqa $xt0,0x00(%rsp)
1080 movdqa $xb3,0x10(%rsp)
1081 lea 0x40($out),$out # out+=64*3
1082 movdqa $xc3,0x20(%rsp)
1083 sub \$192,$len # len-=64*3
1084 movdqa $xd3,0x30(%rsp)
1085
1086 .Loop_tail4x:
1087 movzb ($inp,%r10),%eax
1088 movzb (%rsp,%r10),%ecx
1089 lea 1(%r10),%r10
1090 xor %ecx,%eax
1091 mov %al,-1($out,%r10)
1092 dec $len
1093 jnz .Loop_tail4x
1094
1095 .Ldone4x:
1096 ___
1097 $code.=<<___ if ($win64);
1098 lea 0x140+0x30(%rsp),%r11
1099 movaps -0x30(%r11),%xmm6
1100 movaps -0x20(%r11),%xmm7
1101 movaps -0x10(%r11),%xmm8
1102 movaps 0x00(%r11),%xmm9
1103 movaps 0x10(%r11),%xmm10
1104 movaps 0x20(%r11),%xmm11
1105 movaps 0x30(%r11),%xmm12
1106 movaps 0x40(%r11),%xmm13
1107 movaps 0x50(%r11),%xmm14
1108 movaps 0x60(%r11),%xmm15
1109 ___
1110 $code.=<<___;
1111 add \$0x148+$xframe,%rsp
1112 ret
1113 .size ChaCha20_4x,.-ChaCha20_4x
1114 ___
1115 }
1116
1117 ########################################################################
1118 # AVX2 code path
1119 if ($avx>1) {
1120 my ($xb0,$xb1,$xb2,$xb3, $xd0,$xd1,$xd2,$xd3,
1121 $xa0,$xa1,$xa2,$xa3, $xt0,$xt1,$xt2,$xt3)=map("%ymm$_",(0..15));
1122 my @xx=($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
1123 "%nox","%nox","%nox","%nox", $xd0,$xd1,$xd2,$xd3);
1124
1125 sub AVX2_lane_ROUND {
1126 my ($a0,$b0,$c0,$d0)=@_;
1127 my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
1128 my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
1129 my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
1130 my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
1131 my @x=map("\"$_\"",@xx);
1132
1133 # Consider order in which variables are addressed by their
1134 # index:
1135 #
1136 # a b c d
1137 #
1138 # 0 4 8 12 < even round
1139 # 1 5 9 13
1140 # 2 6 10 14
1141 # 3 7 11 15
1142 # 0 5 10 15 < odd round
1143 # 1 6 11 12
1144 # 2 7 8 13
1145 # 3 4 9 14
1146 #
1147 # 'a', 'b' and 'd's are permanently allocated in registers,
1148 # @x[0..7,12..15], while 'c's are maintained in memory. If
1149 # you observe 'c' column, you'll notice that pair of 'c's is
1150 # invariant between rounds. This means that we have to reload
1151 # them once per round, in the middle. This is why you'll see
1152 # bunch of 'c' stores and loads in the middle, but none in
1153 # the beginning or end.
1154
1155 (
1156 "&vpaddd (@x[$a0],@x[$a0],@x[$b0])", # Q1
1157 "&vpxor (@x[$d0],@x[$a0],@x[$d0])",
1158 "&vpshufb (@x[$d0],@x[$d0],$t1)",
1159 "&vpaddd (@x[$a1],@x[$a1],@x[$b1])", # Q2
1160 "&vpxor (@x[$d1],@x[$a1],@x[$d1])",
1161 "&vpshufb (@x[$d1],@x[$d1],$t1)",
1162
1163 "&vpaddd ($xc,$xc,@x[$d0])",
1164 "&vpxor (@x[$b0],$xc,@x[$b0])",
1165 "&vpslld ($t0,@x[$b0],12)",
1166 "&vpsrld (@x[$b0],@x[$b0],20)",
1167 "&vpor (@x[$b0],$t0,@x[$b0])",
1168 "&vbroadcasti128($t0,'(%r11)')", # .Lrot24(%rip)
1169 "&vpaddd ($xc_,$xc_,@x[$d1])",
1170 "&vpxor (@x[$b1],$xc_,@x[$b1])",
1171 "&vpslld ($t1,@x[$b1],12)",
1172 "&vpsrld (@x[$b1],@x[$b1],20)",
1173 "&vpor (@x[$b1],$t1,@x[$b1])",
1174
1175 "&vpaddd (@x[$a0],@x[$a0],@x[$b0])",
1176 "&vpxor (@x[$d0],@x[$a0],@x[$d0])",
1177 "&vpshufb (@x[$d0],@x[$d0],$t0)",
1178 "&vpaddd (@x[$a1],@x[$a1],@x[$b1])",
1179 "&vpxor (@x[$d1],@x[$a1],@x[$d1])",
1180 "&vpshufb (@x[$d1],@x[$d1],$t0)",
1181
1182 "&vpaddd ($xc,$xc,@x[$d0])",
1183 "&vpxor (@x[$b0],$xc,@x[$b0])",
1184 "&vpslld ($t1,@x[$b0],7)",
1185 "&vpsrld (@x[$b0],@x[$b0],25)",
1186 "&vpor (@x[$b0],$t1,@x[$b0])",
1187 "&vbroadcasti128($t1,'(%r10)')", # .Lrot16(%rip)
1188 "&vpaddd ($xc_,$xc_,@x[$d1])",
1189 "&vpxor (@x[$b1],$xc_,@x[$b1])",
1190 "&vpslld ($t0,@x[$b1],7)",
1191 "&vpsrld (@x[$b1],@x[$b1],25)",
1192 "&vpor (@x[$b1],$t0,@x[$b1])",
1193
1194 "&vmovdqa (\"`32*($c0-8)`(%rsp)\",$xc)", # reload pair of 'c's
1195 "&vmovdqa (\"`32*($c1-8)`(%rsp)\",$xc_)",
1196 "&vmovdqa ($xc,\"`32*($c2-8)`(%rsp)\")",
1197 "&vmovdqa ($xc_,\"`32*($c3-8)`(%rsp)\")",
1198
1199 "&vpaddd (@x[$a2],@x[$a2],@x[$b2])", # Q3
1200 "&vpxor (@x[$d2],@x[$a2],@x[$d2])",
1201 "&vpshufb (@x[$d2],@x[$d2],$t1)",
1202 "&vpaddd (@x[$a3],@x[$a3],@x[$b3])", # Q4
1203 "&vpxor (@x[$d3],@x[$a3],@x[$d3])",
1204 "&vpshufb (@x[$d3],@x[$d3],$t1)",
1205
1206 "&vpaddd ($xc,$xc,@x[$d2])",
1207 "&vpxor (@x[$b2],$xc,@x[$b2])",
1208 "&vpslld ($t0,@x[$b2],12)",
1209 "&vpsrld (@x[$b2],@x[$b2],20)",
1210 "&vpor (@x[$b2],$t0,@x[$b2])",
1211 "&vbroadcasti128($t0,'(%r11)')", # .Lrot24(%rip)
1212 "&vpaddd ($xc_,$xc_,@x[$d3])",
1213 "&vpxor (@x[$b3],$xc_,@x[$b3])",
1214 "&vpslld ($t1,@x[$b3],12)",
1215 "&vpsrld (@x[$b3],@x[$b3],20)",
1216 "&vpor (@x[$b3],$t1,@x[$b3])",
1217
1218 "&vpaddd (@x[$a2],@x[$a2],@x[$b2])",
1219 "&vpxor (@x[$d2],@x[$a2],@x[$d2])",
1220 "&vpshufb (@x[$d2],@x[$d2],$t0)",
1221 "&vpaddd (@x[$a3],@x[$a3],@x[$b3])",
1222 "&vpxor (@x[$d3],@x[$a3],@x[$d3])",
1223 "&vpshufb (@x[$d3],@x[$d3],$t0)",
1224
1225 "&vpaddd ($xc,$xc,@x[$d2])",
1226 "&vpxor (@x[$b2],$xc,@x[$b2])",
1227 "&vpslld ($t1,@x[$b2],7)",
1228 "&vpsrld (@x[$b2],@x[$b2],25)",
1229 "&vpor (@x[$b2],$t1,@x[$b2])",
1230 "&vbroadcasti128($t1,'(%r10)')", # .Lrot16(%rip)
1231 "&vpaddd ($xc_,$xc_,@x[$d3])",
1232 "&vpxor (@x[$b3],$xc_,@x[$b3])",
1233 "&vpslld ($t0,@x[$b3],7)",
1234 "&vpsrld (@x[$b3],@x[$b3],25)",
1235 "&vpor (@x[$b3],$t0,@x[$b3])"
1236 );
1237 }
1238
1239 my $xframe = $win64 ? 0xb0 : 8;
1240
1241 $code.=<<___;
1242 .type ChaCha20_8x,\@function,5
1243 .align 32
1244 ChaCha20_8x:
1245 .LChaCha20_8x:
1246 mov %rsp,%r10
1247 sub \$0x280+$xframe,%rsp
1248 and \$-32,%rsp
1249 ___
1250 $code.=<<___ if ($win64);
1251 lea 0x290+0x30(%rsp),%r11
1252 movaps %xmm6,-0x30(%r11)
1253 movaps %xmm7,-0x20(%r11)
1254 movaps %xmm8,-0x10(%r11)
1255 movaps %xmm9,0x00(%r11)
1256 movaps %xmm10,0x10(%r11)
1257 movaps %xmm11,0x20(%r11)
1258 movaps %xmm12,0x30(%r11)
1259 movaps %xmm13,0x40(%r11)
1260 movaps %xmm14,0x50(%r11)
1261 movaps %xmm15,0x60(%r11)
1262 ___
1263 $code.=<<___;
1264 vzeroupper
1265 mov %r10,0x280(%rsp)
1266
1267 ################ stack layout
1268 # +0x00 SIMD equivalent of @x[8-12]
1269 # ...
1270 # +0x80 constant copy of key[0-2] smashed by lanes
1271 # ...
1272 # +0x200 SIMD counters (with nonce smashed by lanes)
1273 # ...
1274 # +0x280 saved %rsp
1275
1276 vbroadcasti128 .Lsigma(%rip),$xa3 # key[0]
1277 vbroadcasti128 ($key),$xb3 # key[1]
1278 vbroadcasti128 16($key),$xt3 # key[2]
1279 vbroadcasti128 ($counter),$xd3 # key[3]
1280 lea 0x100(%rsp),%rcx # size optimization
1281 lea 0x200(%rsp),%rax # size optimization
1282 lea .Lrot16(%rip),%r10
1283 lea .Lrot24(%rip),%r11
1284
1285 vpshufd \$0x00,$xa3,$xa0 # smash key by lanes...
1286 vpshufd \$0x55,$xa3,$xa1
1287 vmovdqa $xa0,0x80-0x100(%rcx) # ... and offload
1288 vpshufd \$0xaa,$xa3,$xa2
1289 vmovdqa $xa1,0xa0-0x100(%rcx)
1290 vpshufd \$0xff,$xa3,$xa3
1291 vmovdqa $xa2,0xc0-0x100(%rcx)
1292 vmovdqa $xa3,0xe0-0x100(%rcx)
1293
1294 vpshufd \$0x00,$xb3,$xb0
1295 vpshufd \$0x55,$xb3,$xb1
1296 vmovdqa $xb0,0x100-0x100(%rcx)
1297 vpshufd \$0xaa,$xb3,$xb2
1298 vmovdqa $xb1,0x120-0x100(%rcx)
1299 vpshufd \$0xff,$xb3,$xb3
1300 vmovdqa $xb2,0x140-0x100(%rcx)
1301 vmovdqa $xb3,0x160-0x100(%rcx)
1302
1303 vpshufd \$0x00,$xt3,$xt0 # "xc0"
1304 vpshufd \$0x55,$xt3,$xt1 # "xc1"
1305 vmovdqa $xt0,0x180-0x200(%rax)
1306 vpshufd \$0xaa,$xt3,$xt2 # "xc2"
1307 vmovdqa $xt1,0x1a0-0x200(%rax)
1308 vpshufd \$0xff,$xt3,$xt3 # "xc3"
1309 vmovdqa $xt2,0x1c0-0x200(%rax)
1310 vmovdqa $xt3,0x1e0-0x200(%rax)
1311
1312 vpshufd \$0x00,$xd3,$xd0
1313 vpshufd \$0x55,$xd3,$xd1
1314 vpaddd .Lincy(%rip),$xd0,$xd0 # don't save counters yet
1315 vpshufd \$0xaa,$xd3,$xd2
1316 vmovdqa $xd1,0x220-0x200(%rax)
1317 vpshufd \$0xff,$xd3,$xd3
1318 vmovdqa $xd2,0x240-0x200(%rax)
1319 vmovdqa $xd3,0x260-0x200(%rax)
1320
1321 jmp .Loop_enter8x
1322
1323 .align 32
1324 .Loop_outer8x:
1325 vmovdqa 0x80-0x100(%rcx),$xa0 # re-load smashed key
1326 vmovdqa 0xa0-0x100(%rcx),$xa1
1327 vmovdqa 0xc0-0x100(%rcx),$xa2
1328 vmovdqa 0xe0-0x100(%rcx),$xa3
1329 vmovdqa 0x100-0x100(%rcx),$xb0
1330 vmovdqa 0x120-0x100(%rcx),$xb1
1331 vmovdqa 0x140-0x100(%rcx),$xb2
1332 vmovdqa 0x160-0x100(%rcx),$xb3
1333 vmovdqa 0x180-0x200(%rax),$xt0 # "xc0"
1334 vmovdqa 0x1a0-0x200(%rax),$xt1 # "xc1"
1335 vmovdqa 0x1c0-0x200(%rax),$xt2 # "xc2"
1336 vmovdqa 0x1e0-0x200(%rax),$xt3 # "xc3"
1337 vmovdqa 0x200-0x200(%rax),$xd0
1338 vmovdqa 0x220-0x200(%rax),$xd1
1339 vmovdqa 0x240-0x200(%rax),$xd2
1340 vmovdqa 0x260-0x200(%rax),$xd3
1341 vpaddd .Leight(%rip),$xd0,$xd0 # next SIMD counters
1342
1343 .Loop_enter8x:
1344 vmovdqa $xt2,0x40(%rsp) # SIMD equivalent of "@x[10]"
1345 vmovdqa $xt3,0x60(%rsp) # SIMD equivalent of "@x[11]"
1346 vbroadcasti128 (%r10),$xt3
1347 vmovdqa $xd0,0x200-0x200(%rax) # save SIMD counters
1348 mov \$10,%eax
1349 jmp .Loop8x
1350
1351 .align 32
1352 .Loop8x:
1353 ___
1354 foreach (&AVX2_lane_ROUND(0, 4, 8,12)) { eval; }
1355 foreach (&AVX2_lane_ROUND(0, 5,10,15)) { eval; }
1356 $code.=<<___;
1357 dec %eax
1358 jnz .Loop8x
1359
1360 lea 0x200(%rsp),%rax # size optimization
1361 vpaddd 0x80-0x100(%rcx),$xa0,$xa0 # accumulate key
1362 vpaddd 0xa0-0x100(%rcx),$xa1,$xa1
1363 vpaddd 0xc0-0x100(%rcx),$xa2,$xa2
1364 vpaddd 0xe0-0x100(%rcx),$xa3,$xa3
1365
1366 vpunpckldq $xa1,$xa0,$xt2 # "de-interlace" data
1367 vpunpckldq $xa3,$xa2,$xt3
1368 vpunpckhdq $xa1,$xa0,$xa0
1369 vpunpckhdq $xa3,$xa2,$xa2
1370 vpunpcklqdq $xt3,$xt2,$xa1 # "a0"
1371 vpunpckhqdq $xt3,$xt2,$xt2 # "a1"
1372 vpunpcklqdq $xa2,$xa0,$xa3 # "a2"
1373 vpunpckhqdq $xa2,$xa0,$xa0 # "a3"
1374 ___
1375 ($xa0,$xa1,$xa2,$xa3,$xt2)=($xa1,$xt2,$xa3,$xa0,$xa2);
1376 $code.=<<___;
1377 vpaddd 0x100-0x100(%rcx),$xb0,$xb0
1378 vpaddd 0x120-0x100(%rcx),$xb1,$xb1
1379 vpaddd 0x140-0x100(%rcx),$xb2,$xb2
1380 vpaddd 0x160-0x100(%rcx),$xb3,$xb3
1381
1382 vpunpckldq $xb1,$xb0,$xt2
1383 vpunpckldq $xb3,$xb2,$xt3
1384 vpunpckhdq $xb1,$xb0,$xb0
1385 vpunpckhdq $xb3,$xb2,$xb2
1386 vpunpcklqdq $xt3,$xt2,$xb1 # "b0"
1387 vpunpckhqdq $xt3,$xt2,$xt2 # "b1"
1388 vpunpcklqdq $xb2,$xb0,$xb3 # "b2"
1389 vpunpckhqdq $xb2,$xb0,$xb0 # "b3"
1390 ___
1391 ($xb0,$xb1,$xb2,$xb3,$xt2)=($xb1,$xt2,$xb3,$xb0,$xb2);
1392 $code.=<<___;
1393 vperm2i128 \$0x20,$xb0,$xa0,$xt3 # "de-interlace" further
1394 vperm2i128 \$0x31,$xb0,$xa0,$xb0
1395 vperm2i128 \$0x20,$xb1,$xa1,$xa0
1396 vperm2i128 \$0x31,$xb1,$xa1,$xb1
1397 vperm2i128 \$0x20,$xb2,$xa2,$xa1
1398 vperm2i128 \$0x31,$xb2,$xa2,$xb2
1399 vperm2i128 \$0x20,$xb3,$xa3,$xa2
1400 vperm2i128 \$0x31,$xb3,$xa3,$xb3
1401 ___
1402 ($xa0,$xa1,$xa2,$xa3,$xt3)=($xt3,$xa0,$xa1,$xa2,$xa3);
1403 my ($xc0,$xc1,$xc2,$xc3)=($xt0,$xt1,$xa0,$xa1);
1404 $code.=<<___;
1405 vmovdqa $xa0,0x00(%rsp) # offload $xaN
1406 vmovdqa $xa1,0x20(%rsp)
1407 vmovdqa 0x40(%rsp),$xc2 # $xa0
1408 vmovdqa 0x60(%rsp),$xc3 # $xa1
1409
1410 vpaddd 0x180-0x200(%rax),$xc0,$xc0
1411 vpaddd 0x1a0-0x200(%rax),$xc1,$xc1
1412 vpaddd 0x1c0-0x200(%rax),$xc2,$xc2
1413 vpaddd 0x1e0-0x200(%rax),$xc3,$xc3
1414
1415 vpunpckldq $xc1,$xc0,$xt2
1416 vpunpckldq $xc3,$xc2,$xt3
1417 vpunpckhdq $xc1,$xc0,$xc0
1418 vpunpckhdq $xc3,$xc2,$xc2
1419 vpunpcklqdq $xt3,$xt2,$xc1 # "c0"
1420 vpunpckhqdq $xt3,$xt2,$xt2 # "c1"
1421 vpunpcklqdq $xc2,$xc0,$xc3 # "c2"
1422 vpunpckhqdq $xc2,$xc0,$xc0 # "c3"
1423 ___
1424 ($xc0,$xc1,$xc2,$xc3,$xt2)=($xc1,$xt2,$xc3,$xc0,$xc2);
1425 $code.=<<___;
1426 vpaddd 0x200-0x200(%rax),$xd0,$xd0
1427 vpaddd 0x220-0x200(%rax),$xd1,$xd1
1428 vpaddd 0x240-0x200(%rax),$xd2,$xd2
1429 vpaddd 0x260-0x200(%rax),$xd3,$xd3
1430
1431 vpunpckldq $xd1,$xd0,$xt2
1432 vpunpckldq $xd3,$xd2,$xt3
1433 vpunpckhdq $xd1,$xd0,$xd0
1434 vpunpckhdq $xd3,$xd2,$xd2
1435 vpunpcklqdq $xt3,$xt2,$xd1 # "d0"
1436 vpunpckhqdq $xt3,$xt2,$xt2 # "d1"
1437 vpunpcklqdq $xd2,$xd0,$xd3 # "d2"
1438 vpunpckhqdq $xd2,$xd0,$xd0 # "d3"
1439 ___
1440 ($xd0,$xd1,$xd2,$xd3,$xt2)=($xd1,$xt2,$xd3,$xd0,$xd2);
1441 $code.=<<___;
1442 vperm2i128 \$0x20,$xd0,$xc0,$xt3 # "de-interlace" further
1443 vperm2i128 \$0x31,$xd0,$xc0,$xd0
1444 vperm2i128 \$0x20,$xd1,$xc1,$xc0
1445 vperm2i128 \$0x31,$xd1,$xc1,$xd1
1446 vperm2i128 \$0x20,$xd2,$xc2,$xc1
1447 vperm2i128 \$0x31,$xd2,$xc2,$xd2
1448 vperm2i128 \$0x20,$xd3,$xc3,$xc2
1449 vperm2i128 \$0x31,$xd3,$xc3,$xd3
1450 ___
1451 ($xc0,$xc1,$xc2,$xc3,$xt3)=($xt3,$xc0,$xc1,$xc2,$xc3);
1452 ($xb0,$xb1,$xb2,$xb3,$xc0,$xc1,$xc2,$xc3)=
1453 ($xc0,$xc1,$xc2,$xc3,$xb0,$xb1,$xb2,$xb3);
1454 ($xa0,$xa1)=($xt2,$xt3);
1455 $code.=<<___;
1456 vmovdqa 0x00(%rsp),$xa0 # $xaN was offloaded, remember?
1457 vmovdqa 0x20(%rsp),$xa1
1458
1459 cmp \$64*8,$len
1460 jb .Ltail8x
1461
1462 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1463 vpxor 0x20($inp),$xb0,$xb0
1464 vpxor 0x40($inp),$xc0,$xc0
1465 vpxor 0x60($inp),$xd0,$xd0
1466 lea 0x80($inp),$inp # size optimization
1467 vmovdqu $xa0,0x00($out)
1468 vmovdqu $xb0,0x20($out)
1469 vmovdqu $xc0,0x40($out)
1470 vmovdqu $xd0,0x60($out)
1471 lea 0x80($out),$out # size optimization
1472
1473 vpxor 0x00($inp),$xa1,$xa1
1474 vpxor 0x20($inp),$xb1,$xb1
1475 vpxor 0x40($inp),$xc1,$xc1
1476 vpxor 0x60($inp),$xd1,$xd1
1477 lea 0x80($inp),$inp # size optimization
1478 vmovdqu $xa1,0x00($out)
1479 vmovdqu $xb1,0x20($out)
1480 vmovdqu $xc1,0x40($out)
1481 vmovdqu $xd1,0x60($out)
1482 lea 0x80($out),$out # size optimization
1483
1484 vpxor 0x00($inp),$xa2,$xa2
1485 vpxor 0x20($inp),$xb2,$xb2
1486 vpxor 0x40($inp),$xc2,$xc2
1487 vpxor 0x60($inp),$xd2,$xd2
1488 lea 0x80($inp),$inp # size optimization
1489 vmovdqu $xa2,0x00($out)
1490 vmovdqu $xb2,0x20($out)
1491 vmovdqu $xc2,0x40($out)
1492 vmovdqu $xd2,0x60($out)
1493 lea 0x80($out),$out # size optimization
1494
1495 vpxor 0x00($inp),$xa3,$xa3
1496 vpxor 0x20($inp),$xb3,$xb3
1497 vpxor 0x40($inp),$xc3,$xc3
1498 vpxor 0x60($inp),$xd3,$xd3
1499 lea 0x80($inp),$inp # size optimization
1500 vmovdqu $xa3,0x00($out)
1501 vmovdqu $xb3,0x20($out)
1502 vmovdqu $xc3,0x40($out)
1503 vmovdqu $xd3,0x60($out)
1504 lea 0x80($out),$out # size optimization
1505
1506 sub \$64*8,$len
1507 jnz .Loop_outer8x
1508
1509 jmp .Ldone8x
1510
1511 .Ltail8x:
1512 cmp \$448,$len
1513 jae .L448_or_more8x
1514 cmp \$384,$len
1515 jae .L384_or_more8x
1516 cmp \$320,$len
1517 jae .L320_or_more8x
1518 cmp \$256,$len
1519 jae .L256_or_more8x
1520 cmp \$192,$len
1521 jae .L192_or_more8x
1522 cmp \$128,$len
1523 jae .L128_or_more8x
1524 cmp \$64,$len
1525 jae .L64_or_more8x
1526
1527 xor %r10,%r10
1528 vmovdqa $xa0,0x00(%rsp)
1529 vmovdqa $xb0,0x20(%rsp)
1530 jmp .Loop_tail8x
1531
1532 .align 32
1533 .L64_or_more8x:
1534 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1535 vpxor 0x20($inp),$xb0,$xb0
1536 vmovdqu $xa0,0x00($out)
1537 vmovdqu $xb0,0x20($out)
1538 je .Ldone8x
1539
1540 lea 0x40($inp),$inp # inp+=64*1
1541 xor %r10,%r10
1542 vmovdqa $xc0,0x00(%rsp)
1543 lea 0x40($out),$out # out+=64*1
1544 sub \$64,$len # len-=64*1
1545 vmovdqa $xd0,0x20(%rsp)
1546 jmp .Loop_tail8x
1547
1548 .align 32
1549 .L128_or_more8x:
1550 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1551 vpxor 0x20($inp),$xb0,$xb0
1552 vpxor 0x40($inp),$xc0,$xc0
1553 vpxor 0x60($inp),$xd0,$xd0
1554 vmovdqu $xa0,0x00($out)
1555 vmovdqu $xb0,0x20($out)
1556 vmovdqu $xc0,0x40($out)
1557 vmovdqu $xd0,0x60($out)
1558 je .Ldone8x
1559
1560 lea 0x80($inp),$inp # inp+=64*2
1561 xor %r10,%r10
1562 vmovdqa $xa1,0x00(%rsp)
1563 lea 0x80($out),$out # out+=64*2
1564 sub \$128,$len # len-=64*2
1565 vmovdqa $xb1,0x20(%rsp)
1566 jmp .Loop_tail8x
1567
1568 .align 32
1569 .L192_or_more8x:
1570 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1571 vpxor 0x20($inp),$xb0,$xb0
1572 vpxor 0x40($inp),$xc0,$xc0
1573 vpxor 0x60($inp),$xd0,$xd0
1574 vpxor 0x80($inp),$xa1,$xa1
1575 vpxor 0xa0($inp),$xb1,$xb1
1576 vmovdqu $xa0,0x00($out)
1577 vmovdqu $xb0,0x20($out)
1578 vmovdqu $xc0,0x40($out)
1579 vmovdqu $xd0,0x60($out)
1580 vmovdqu $xa1,0x80($out)
1581 vmovdqu $xb1,0xa0($out)
1582 je .Ldone8x
1583
1584 lea 0xc0($inp),$inp # inp+=64*3
1585 xor %r10,%r10
1586 vmovdqa $xc1,0x00(%rsp)
1587 lea 0xc0($out),$out # out+=64*3
1588 sub \$192,$len # len-=64*3
1589 vmovdqa $xd1,0x20(%rsp)
1590 jmp .Loop_tail8x
1591
1592 .align 32
1593 .L256_or_more8x:
1594 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1595 vpxor 0x20($inp),$xb0,$xb0
1596 vpxor 0x40($inp),$xc0,$xc0
1597 vpxor 0x60($inp),$xd0,$xd0
1598 vpxor 0x80($inp),$xa1,$xa1
1599 vpxor 0xa0($inp),$xb1,$xb1
1600 vpxor 0xc0($inp),$xc1,$xc1
1601 vpxor 0xe0($inp),$xd1,$xd1
1602 vmovdqu $xa0,0x00($out)
1603 vmovdqu $xb0,0x20($out)
1604 vmovdqu $xc0,0x40($out)
1605 vmovdqu $xd0,0x60($out)
1606 vmovdqu $xa1,0x80($out)
1607 vmovdqu $xb1,0xa0($out)
1608 vmovdqu $xc1,0xc0($out)
1609 vmovdqu $xd1,0xe0($out)
1610 je .Ldone8x
1611
1612 lea 0x100($inp),$inp # inp+=64*4
1613 xor %r10,%r10
1614 vmovdqa $xa2,0x00(%rsp)
1615 lea 0x100($out),$out # out+=64*4
1616 sub \$256,$len # len-=64*4
1617 vmovdqa $xb2,0x20(%rsp)
1618 jmp .Loop_tail8x
1619
1620 .align 32
1621 .L320_or_more8x:
1622 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1623 vpxor 0x20($inp),$xb0,$xb0
1624 vpxor 0x40($inp),$xc0,$xc0
1625 vpxor 0x60($inp),$xd0,$xd0
1626 vpxor 0x80($inp),$xa1,$xa1
1627 vpxor 0xa0($inp),$xb1,$xb1
1628 vpxor 0xc0($inp),$xc1,$xc1
1629 vpxor 0xe0($inp),$xd1,$xd1
1630 vpxor 0x100($inp),$xa2,$xa2
1631 vpxor 0x120($inp),$xb2,$xb2
1632 vmovdqu $xa0,0x00($out)
1633 vmovdqu $xb0,0x20($out)
1634 vmovdqu $xc0,0x40($out)
1635 vmovdqu $xd0,0x60($out)
1636 vmovdqu $xa1,0x80($out)
1637 vmovdqu $xb1,0xa0($out)
1638 vmovdqu $xc1,0xc0($out)
1639 vmovdqu $xd1,0xe0($out)
1640 vmovdqu $xa2,0x100($out)
1641 vmovdqu $xb2,0x120($out)
1642 je .Ldone8x
1643
1644 lea 0x140($inp),$inp # inp+=64*5
1645 xor %r10,%r10
1646 vmovdqa $xc2,0x00(%rsp)
1647 lea 0x140($out),$out # out+=64*5
1648 sub \$320,$len # len-=64*5
1649 vmovdqa $xd2,0x20(%rsp)
1650 jmp .Loop_tail8x
1651
1652 .align 32
1653 .L384_or_more8x:
1654 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1655 vpxor 0x20($inp),$xb0,$xb0
1656 vpxor 0x40($inp),$xc0,$xc0
1657 vpxor 0x60($inp),$xd0,$xd0
1658 vpxor 0x80($inp),$xa1,$xa1
1659 vpxor 0xa0($inp),$xb1,$xb1
1660 vpxor 0xc0($inp),$xc1,$xc1
1661 vpxor 0xe0($inp),$xd1,$xd1
1662 vpxor 0x100($inp),$xa2,$xa2
1663 vpxor 0x120($inp),$xb2,$xb2
1664 vpxor 0x140($inp),$xc2,$xc2
1665 vpxor 0x160($inp),$xd2,$xd2
1666 vmovdqu $xa0,0x00($out)
1667 vmovdqu $xb0,0x20($out)
1668 vmovdqu $xc0,0x40($out)
1669 vmovdqu $xd0,0x60($out)
1670 vmovdqu $xa1,0x80($out)
1671 vmovdqu $xb1,0xa0($out)
1672 vmovdqu $xc1,0xc0($out)
1673 vmovdqu $xd1,0xe0($out)
1674 vmovdqu $xa2,0x100($out)
1675 vmovdqu $xb2,0x120($out)
1676 vmovdqu $xc2,0x140($out)
1677 vmovdqu $xd2,0x160($out)
1678 je .Ldone8x
1679
1680 lea 0x180($inp),$inp # inp+=64*6
1681 xor %r10,%r10
1682 vmovdqa $xa3,0x00(%rsp)
1683 lea 0x180($out),$out # out+=64*6
1684 sub \$384,$len # len-=64*6
1685 vmovdqa $xb3,0x20(%rsp)
1686 jmp .Loop_tail8x
1687
1688 .align 32
1689 .L448_or_more8x:
1690 vpxor 0x00($inp),$xa0,$xa0 # xor with input
1691 vpxor 0x20($inp),$xb0,$xb0
1692 vpxor 0x40($inp),$xc0,$xc0
1693 vpxor 0x60($inp),$xd0,$xd0
1694 vpxor 0x80($inp),$xa1,$xa1
1695 vpxor 0xa0($inp),$xb1,$xb1
1696 vpxor 0xc0($inp),$xc1,$xc1
1697 vpxor 0xe0($inp),$xd1,$xd1
1698 vpxor 0x100($inp),$xa2,$xa2
1699 vpxor 0x120($inp),$xb2,$xb2
1700 vpxor 0x140($inp),$xc2,$xc2
1701 vpxor 0x160($inp),$xd2,$xd2
1702 vpxor 0x180($inp),$xa3,$xa3
1703 vpxor 0x1a0($inp),$xb3,$xb3
1704 vmovdqu $xa0,0x00($out)
1705 vmovdqu $xb0,0x20($out)
1706 vmovdqu $xc0,0x40($out)
1707 vmovdqu $xd0,0x60($out)
1708 vmovdqu $xa1,0x80($out)
1709 vmovdqu $xb1,0xa0($out)
1710 vmovdqu $xc1,0xc0($out)
1711 vmovdqu $xd1,0xe0($out)
1712 vmovdqu $xa2,0x100($out)
1713 vmovdqu $xb2,0x120($out)
1714 vmovdqu $xc2,0x140($out)
1715 vmovdqu $xd2,0x160($out)
1716 vmovdqu $xa3,0x180($out)
1717 vmovdqu $xb3,0x1a0($out)
1718 je .Ldone8x
1719
1720 lea 0x1c0($inp),$inp # inp+=64*7
1721 xor %r10,%r10
1722 vmovdqa $xc3,0x00(%rsp)
1723 lea 0x1c0($out),$out # out+=64*7
1724 sub \$448,$len # len-=64*7
1725 vmovdqa $xd3,0x20(%rsp)
1726
1727 .Loop_tail8x:
1728 movzb ($inp,%r10),%eax
1729 movzb (%rsp,%r10),%ecx
1730 lea 1(%r10),%r10
1731 xor %ecx,%eax
1732 mov %al,-1($out,%r10)
1733 dec $len
1734 jnz .Loop_tail8x
1735
1736 .Ldone8x:
1737 vzeroall
1738 ___
1739 $code.=<<___ if ($win64);
1740 lea 0x290+0x30(%rsp),%r11
1741 movaps -0x30(%r11),%xmm6
1742 movaps -0x20(%r11),%xmm7
1743 movaps -0x10(%r11),%xmm8
1744 movaps 0x00(%r11),%xmm9
1745 movaps 0x10(%r11),%xmm10
1746 movaps 0x20(%r11),%xmm11
1747 movaps 0x30(%r11),%xmm12
1748 movaps 0x40(%r11),%xmm13
1749 movaps 0x50(%r11),%xmm14
1750 movaps 0x60(%r11),%xmm15
1751 ___
1752 $code.=<<___;
1753 mov 0x280(%rsp),%rsp
1754 ret
1755 .size ChaCha20_8x,.-ChaCha20_8x
1756 ___
1757 }
1758
1759 foreach (split("\n",$code)) {
1760 s/\`([^\`]*)\`/eval $1/geo;
1761
1762 s/%x#%y/%x/go;
1763
1764 print $_,"\n";
1765 }
1766
1767 close STDOUT;