Debian Codegraph @

external/boringssl/crypto/poly1305/asm/poly1305-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 # This module implements Poly1305 hash for x86_64.
11 #
12 # March 2015
13 #
14 # Numbers are cycles per processed byte with poly1305_blocks alone,
15 # measured with rdtsc at fixed clock frequency.
16 #
17 # IALU/gcc-4.8(*) AVX(**) AVX2
18 # P4 4.90/+120% -
19 # Core 2 2.39/+90% -
20 # Westmere 1.86/+120% -
21 # Sandy Bridge 1.39/+140% 1.10
22 # Haswell 1.10/+175% 1.11 0.65
23 # Skylake 1.12/+120% 0.96 0.51
24 # Silvermont 2.83/+95% -
25 # VIA Nano 1.82/+150% -
26 # Sledgehammer 1.38/+160% -
27 # Bulldozer 2.21/+130% 0.97
28 #
29 # (*) improvement coefficients relative to clang are more modest and
30 # are ~50% on most processors, in both cases we are comparing to
31 # __int128 code;
32 # (**) SSE2 implementation was attempted, but among non-AVX processors
33 # it was faster than integer-only code only on older Intel P4 and
34 # Core processors, 50-30%, less newer processor is, but slower on
35 # contemporary ones, for example almost 2x slower on Atom, and as
36 # former are naturally disappearing, SSE2 is deemed unnecessary;
37
38 $flavour = shift;
39 $output = shift;
40 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
41
42 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
43
44 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
45 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
46 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
47 die "can't locate x86_64-xlate.pl";
48
49 $avx = 2;
50
51 open OUT,"| \"$^X\" $xlate $flavour $output";
52 *STDOUT=*OUT;
53
54 my ($ctx,$inp,$len,$padbit)=("%rdi","%rsi","%rdx","%rcx");
55 my ($mac,$nonce)=($inp,$len); # *_emit arguments
56 my ($d1,$d2,$d3, $r0,$r1,$s1)=map("%r$_",(8..13));
57 my ($h0,$h1,$h2)=("%r14","%rbx","%rbp");
58
59 sub poly1305_iteration {
60 # input: copy of $r1 in %rax, $h0-$h2, $r0-$r1
61 # output: $h0-$h2 *= $r0-$r1
62 $code.=<<___;
63 mulq $h0 # h0*r1
64 mov %rax,$d2
65 mov $r0,%rax
66 mov %rdx,$d3
67
68 mulq $h0 # h0*r0
69 mov %rax,$h0 # future $h0
70 mov $r0,%rax
71 mov %rdx,$d1
72
73 mulq $h1 # h1*r0
74 add %rax,$d2
75 mov $s1,%rax
76 adc %rdx,$d3
77
78 mulq $h1 # h1*s1
79 mov $h2,$h1 # borrow $h1
80 add %rax,$h0
81 adc %rdx,$d1
82
83 imulq $s1,$h1 # h2*s1
84 add $h1,$d2
85 mov $d1,$h1
86 adc \$0,$d3
87
88 imulq $r0,$h2 # h2*r0
89 add $d2,$h1
90 mov \$-4,%rax # mask value
91 adc $h2,$d3
92
93 and $d3,%rax # last reduction step
94 mov $d3,$h2
95 shr \$2,$d3
96 and \$3,$h2
97 add $d3,%rax
98 add %rax,$h0
99 adc \$0,$h1
100 ___
101 }
102
103 ########################################################################
104 # Layout of opaque area is following.
105 #
106 # unsigned __int64 h[3]; # current hash value base 2^64
107 # unsigned __int64 r[2]; # key value base 2^64
108
109 $code.=<<___;
110 .text
111
112 .extern OPENSSL_ia32cap_P
113
114 .globl poly1305_init
115 .globl poly1305_blocks
116 .globl poly1305_emit
117 .type poly1305_init,\@function,3
118 .align 32
119 poly1305_init:
120 xor %rax,%rax
121 mov %rax,0($ctx) # initialize hash value
122 mov %rax,8($ctx)
123 mov %rax,16($ctx)
124
125 cmp \$0,$inp
126 je .Lno_key
127
128 lea poly1305_blocks(%rip),%r10
129 lea poly1305_emit(%rip),%r11
130 ___
131 $code.=<<___ if ($avx);
132 mov OPENSSL_ia32cap_P+4(%rip),%r9
133 lea poly1305_blocks_avx(%rip),%rax
134 lea poly1305_emit_avx(%rip),%rcx
135 bt \$`60-32`,%r9 # AVX?
136 cmovc %rax,%r10
137 cmovc %rcx,%r11
138 ___
139 $code.=<<___ if ($avx>1);
140 lea poly1305_blocks_avx2(%rip),%rax
141 bt \$`5+32`,%r9 # AVX2?
142 cmovc %rax,%r10
143 ___
144 $code.=<<___;
145 mov \$0x0ffffffc0fffffff,%rax
146 mov \$0x0ffffffc0ffffffc,%rcx
147 and 0($inp),%rax
148 and 8($inp),%rcx
149 mov %rax,24($ctx)
150 mov %rcx,32($ctx)
151 ___
152 $code.=<<___ if ($flavour !~ /elf32/);
153 mov %r10,0(%rdx)
154 mov %r11,8(%rdx)
155 ___
156 $code.=<<___ if ($flavour =~ /elf32/);
157 mov %r10d,0(%rdx)
158 mov %r11d,4(%rdx)
159 ___
160 $code.=<<___;
161 mov \$1,%eax
162 .Lno_key:
163 ret
164 .size poly1305_init,.-poly1305_init
165
166 .type poly1305_blocks,\@function,4
167 .align 32
168 poly1305_blocks:
169 .Lblocks:
170 sub \$16,$len # too short?
171 jc .Lno_data
172
173 push %rbx
174 push %rbp
175 push %r12
176 push %r13
177 push %r14
178 push %r15
179 .Lblocks_body:
180
181 mov $len,%r15 # reassign $len
182
183 mov 24($ctx),$r0 # load r
184 mov 32($ctx),$s1
185
186 mov 0($ctx),$h0 # load hash value
187 mov 8($ctx),$h1
188 mov 16($ctx),$h2
189
190 mov $s1,$r1
191 shr \$2,$s1
192 mov $r1,%rax
193 add $r1,$s1 # s1 = r1 + (r1 >> 2)
194 jmp .Loop
195
196 .align 32
197 .Loop:
198 add 0($inp),$h0 # accumulate input
199 adc 8($inp),$h1
200 lea 16($inp),$inp
201 adc $padbit,$h2
202 ___
203 &poly1305_iteration();
204 $code.=<<___;
205 mov $r1,%rax
206 sub \$16,%r15 # len-=16
207 jnc .Loop
208
209 mov $h0,0($ctx) # store hash value
210 mov $h1,8($ctx)
211 mov $h2,16($ctx)
212
213 mov 0(%rsp),%r15
214 mov 8(%rsp),%r14
215 mov 16(%rsp),%r13
216 mov 24(%rsp),%r12
217 mov 32(%rsp),%rbp
218 mov 40(%rsp),%rbx
219 lea 48(%rsp),%rsp
220 .Lno_data:
221 .Lblocks_epilogue:
222 ret
223 .size poly1305_blocks,.-poly1305_blocks
224
225 .type poly1305_emit,\@function,3
226 .align 32
227 poly1305_emit:
228 .Lemit:
229 mov 0($ctx),%r8 # load hash value
230 mov 8($ctx),%r9
231 mov 16($ctx),%r10
232
233 mov %r8,%rax
234 add \$5,%r8 # compare to modulus
235 mov %r9,%rcx
236 adc \$0,%r9
237 adc \$0,%r10
238 shr \$2,%r10 # did 130-bit value overfow?
239 cmovnz %r8,%rax
240 cmovnz %r9,%rcx
241
242 add 0($nonce),%rax # accumulate nonce
243 adc 8($nonce),%rcx
244 mov %rax,0($mac) # write result
245 mov %rcx,8($mac)
246
247 ret
248 .size poly1305_emit,.-poly1305_emit
249 ___
250 if ($avx) {
251
252 ########################################################################
253 # Layout of opaque area is following.
254 #
255 # unsigned __int32 h[5]; # current hash value base 2^26
256 # unsigned __int32 is_base2_26;
257 # unsigned __int64 r[2]; # key value base 2^64
258 # unsigned __int64 pad;
259 # struct { unsigned __int32 r^2, r^1, r^4, r^3; } r[9];
260 #
261 # where r^n are base 2^26 digits of degrees of multiplier key. There are
262 # 5 digits, but last four are interleaved with multiples of 5, totalling
263 # in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
264
265 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
266 map("%xmm$_",(0..15));
267
268 $code.=<<___;
269 .type __poly1305_block,\@abi-omnipotent
270 .align 32
271 __poly1305_block:
272 ___
273 &poly1305_iteration();
274 $code.=<<___;
275 ret
276 .size __poly1305_block,.-__poly1305_block
277
278 .type __poly1305_init_avx,\@abi-omnipotent
279 .align 32
280 __poly1305_init_avx:
281 mov $r0,$h0
282 mov $r1,$h1
283 xor $h2,$h2
284
285 lea 48+64($ctx),$ctx # size optimization
286
287 mov $r1,%rax
288 call __poly1305_block # r^2
289
290 mov \$0x3ffffff,%eax # save interleaved r^2 and r base 2^26
291 mov \$0x3ffffff,%edx
292 mov $h0,$d1
293 and $h0#d,%eax
294 mov $r0,$d2
295 and $r0#d,%edx
296 mov %eax,`16*0+0-64`($ctx)
297 shr \$26,$d1
298 mov %edx,`16*0+4-64`($ctx)
299 shr \$26,$d2
300
301 mov \$0x3ffffff,%eax
302 mov \$0x3ffffff,%edx
303 and $d1#d,%eax
304 and $d2#d,%edx
305 mov %eax,`16*1+0-64`($ctx)
306 lea (%rax,%rax,4),%eax # *5
307 mov %edx,`16*1+4-64`($ctx)
308 lea (%rdx,%rdx,4),%edx # *5
309 mov %eax,`16*2+0-64`($ctx)
310 shr \$26,$d1
311 mov %edx,`16*2+4-64`($ctx)
312 shr \$26,$d2
313
314 mov $h1,%rax
315 mov $r1,%rdx
316 shl \$12,%rax
317 shl \$12,%rdx
318 or $d1,%rax
319 or $d2,%rdx
320 and \$0x3ffffff,%eax
321 and \$0x3ffffff,%edx
322 mov %eax,`16*3+0-64`($ctx)
323 lea (%rax,%rax,4),%eax # *5
324 mov %edx,`16*3+4-64`($ctx)
325 lea (%rdx,%rdx,4),%edx # *5
326 mov %eax,`16*4+0-64`($ctx)
327 mov $h1,$d1
328 mov %edx,`16*4+4-64`($ctx)
329 mov $r1,$d2
330
331 mov \$0x3ffffff,%eax
332 mov \$0x3ffffff,%edx
333 shr \$14,$d1
334 shr \$14,$d2
335 and $d1#d,%eax
336 and $d2#d,%edx
337 mov %eax,`16*5+0-64`($ctx)
338 lea (%rax,%rax,4),%eax # *5
339 mov %edx,`16*5+4-64`($ctx)
340 lea (%rdx,%rdx,4),%edx # *5
341 mov %eax,`16*6+0-64`($ctx)
342 shr \$26,$d1
343 mov %edx,`16*6+4-64`($ctx)
344 shr \$26,$d2
345
346 mov $h2,%rax
347 shl \$24,%rax
348 or %rax,$d1
349 mov $d1#d,`16*7+0-64`($ctx)
350 lea ($d1,$d1,4),$d1 # *5
351 mov $d2#d,`16*7+4-64`($ctx)
352 lea ($d2,$d2,4),$d2 # *5
353 mov $d1#d,`16*8+0-64`($ctx)
354 mov $d2#d,`16*8+4-64`($ctx)
355
356 mov $r1,%rax
357 call __poly1305_block # r^3
358
359 mov \$0x3ffffff,%eax # save r^3 base 2^26
360 mov $h0,$d1
361 and $h0#d,%eax
362 shr \$26,$d1
363 mov %eax,`16*0+12-64`($ctx)
364
365 mov \$0x3ffffff,%edx
366 and $d1#d,%edx
367 mov %edx,`16*1+12-64`($ctx)
368 lea (%rdx,%rdx,4),%edx # *5
369 shr \$26,$d1
370 mov %edx,`16*2+12-64`($ctx)
371
372 mov $h1,%rax
373 shl \$12,%rax
374 or $d1,%rax
375 and \$0x3ffffff,%eax
376 mov %eax,`16*3+12-64`($ctx)
377 lea (%rax,%rax,4),%eax # *5
378 mov $h1,$d1
379 mov %eax,`16*4+12-64`($ctx)
380
381 mov \$0x3ffffff,%edx
382 shr \$14,$d1
383 and $d1#d,%edx
384 mov %edx,`16*5+12-64`($ctx)
385 lea (%rdx,%rdx,4),%edx # *5
386 shr \$26,$d1
387 mov %edx,`16*6+12-64`($ctx)
388
389 mov $h2,%rax
390 shl \$24,%rax
391 or %rax,$d1
392 mov $d1#d,`16*7+12-64`($ctx)
393 lea ($d1,$d1,4),$d1 # *5
394 mov $d1#d,`16*8+12-64`($ctx)
395
396 mov $r1,%rax
397 call __poly1305_block # r^4
398
399 mov \$0x3ffffff,%eax # save r^4 base 2^26
400 mov $h0,$d1
401 and $h0#d,%eax
402 shr \$26,$d1
403 mov %eax,`16*0+8-64`($ctx)
404
405 mov \$0x3ffffff,%edx
406 and $d1#d,%edx
407 mov %edx,`16*1+8-64`($ctx)
408 lea (%rdx,%rdx,4),%edx # *5
409 shr \$26,$d1
410 mov %edx,`16*2+8-64`($ctx)
411
412 mov $h1,%rax
413 shl \$12,%rax
414 or $d1,%rax
415 and \$0x3ffffff,%eax
416 mov %eax,`16*3+8-64`($ctx)
417 lea (%rax,%rax,4),%eax # *5
418 mov $h1,$d1
419 mov %eax,`16*4+8-64`($ctx)
420
421 mov \$0x3ffffff,%edx
422 shr \$14,$d1
423 and $d1#d,%edx
424 mov %edx,`16*5+8-64`($ctx)
425 lea (%rdx,%rdx,4),%edx # *5
426 shr \$26,$d1
427 mov %edx,`16*6+8-64`($ctx)
428
429 mov $h2,%rax
430 shl \$24,%rax
431 or %rax,$d1
432 mov $d1#d,`16*7+8-64`($ctx)
433 lea ($d1,$d1,4),$d1 # *5
434 mov $d1#d,`16*8+8-64`($ctx)
435
436 lea -48-64($ctx),$ctx # size [de-]optimization
437 ret
438 .size __poly1305_init_avx,.-__poly1305_init_avx
439
440 .type poly1305_blocks_avx,\@function,4
441 .align 32
442 poly1305_blocks_avx:
443 mov 20($ctx),%r8d # is_base2_26
444 cmp \$128,$len
445 jae .Lblocks_avx
446 test %r8d,%r8d
447 jz .Lblocks
448
449 .Lblocks_avx:
450 and \$-16,$len
451 jz .Lno_data_avx
452
453 vzeroupper
454
455 test %r8d,%r8d
456 jz .Lbase2_64_avx
457
458 test \$31,$len
459 jz .Leven_avx
460
461 push %rbx
462 push %rbp
463 push %r12
464 push %r13
465 push %r14
466 push %r15
467 .Lblocks_avx_body:
468
469 mov $len,%r15 # reassign $len
470
471 mov 0($ctx),$d1 # load hash value
472 mov 8($ctx),$d2
473 mov 16($ctx),$h2#d
474
475 mov 24($ctx),$r0 # load r
476 mov 32($ctx),$s1
477
478 ################################# base 2^26 -> base 2^64
479 mov $d1#d,$h0#d
480 and \$-1<<31,$d1
481 mov $d2,$r1 # borrow $r1
482 mov $d2#d,$h1#d
483 and \$-1<<31,$d2
484
485 shr \$6,$d1
486 shl \$52,$r1
487 add $d1,$h0
488 shr \$12,$h1
489 shr \$18,$d2
490 add $r1,$h0
491 adc $d2,$h1
492
493 mov $h2,$d1
494 shl \$40,$d1
495 shr \$24,$h2
496 add $d1,$h1
497 adc \$0,$h2 # can be partially reduced...
498
499 mov \$-4,$d2 # ... so reduce
500 mov $h2,$d1
501 and $h2,$d2
502 shr \$2,$d1
503 and \$3,$h2
504 add $d2,$d1 # =*5
505 add $d1,$h0
506 adc \$0,$h1
507
508 mov $s1,$r1
509 mov $s1,%rax
510 shr \$2,$s1
511 add $r1,$s1 # s1 = r1 + (r1 >> 2)
512
513 add 0($inp),$h0 # accumulate input
514 adc 8($inp),$h1
515 lea 16($inp),$inp
516 adc $padbit,$h2
517
518 call __poly1305_block
519
520 test $padbit,$padbit # if $padbit is zero,
521 jz .Lstore_base2_64_avx # store hash in base 2^64 format
522
523 ################################# base 2^64 -> base 2^26
524 mov $h0,%rax
525 mov $h0,%rdx
526 shr \$52,$h0
527 mov $h1,$r0
528 mov $h1,$r1
529 shr \$26,%rdx
530 and \$0x3ffffff,%rax # h[0]
531 shl \$12,$r0
532 and \$0x3ffffff,%rdx # h[1]
533 shr \$14,$h1
534 or $r0,$h0
535 shl \$24,$h2
536 and \$0x3ffffff,$h0 # h[2]
537 shr \$40,$r1
538 and \$0x3ffffff,$h1 # h[3]
539 or $r1,$h2 # h[4]
540
541 sub \$16,%r15
542 jz .Lstore_base2_26_avx
543
544 vmovd %rax#d,$H0
545 vmovd %rdx#d,$H1
546 vmovd $h0#d,$H2
547 vmovd $h1#d,$H3
548 vmovd $h2#d,$H4
549 jmp .Lproceed_avx
550
551 .align 32
552 .Lstore_base2_64_avx:
553 mov $h0,0($ctx)
554 mov $h1,8($ctx)
555 mov $h2,16($ctx) # note that is_base2_26 is zeroed
556 jmp .Ldone_avx
557
558 .align 16
559 .Lstore_base2_26_avx:
560 mov %rax#d,0($ctx) # store hash value base 2^26
561 mov %rdx#d,4($ctx)
562 mov $h0#d,8($ctx)
563 mov $h1#d,12($ctx)
564 mov $h2#d,16($ctx)
565 .align 16
566 .Ldone_avx:
567 mov 0(%rsp),%r15
568 mov 8(%rsp),%r14
569 mov 16(%rsp),%r13
570 mov 24(%rsp),%r12
571 mov 32(%rsp),%rbp
572 mov 40(%rsp),%rbx
573 lea 48(%rsp),%rsp
574 .Lno_data_avx:
575 .Lblocks_avx_epilogue:
576 ret
577
578 .align 32
579 .Lbase2_64_avx:
580 push %rbx
581 push %rbp
582 push %r12
583 push %r13
584 push %r14
585 push %r15
586 .Lbase2_64_avx_body:
587
588 mov $len,%r15 # reassign $len
589
590 mov 24($ctx),$r0 # load r
591 mov 32($ctx),$s1
592
593 mov 0($ctx),$h0 # load hash value
594 mov 8($ctx),$h1
595 mov 16($ctx),$h2#d
596
597 mov $s1,$r1
598 mov $s1,%rax
599 shr \$2,$s1
600 add $r1,$s1 # s1 = r1 + (r1 >> 2)
601
602 test \$31,$len
603 jz .Linit_avx
604
605 add 0($inp),$h0 # accumulate input
606 adc 8($inp),$h1
607 lea 16($inp),$inp
608 adc $padbit,$h2
609 sub \$16,%r15
610
611 call __poly1305_block
612
613 .Linit_avx:
614 ################################# base 2^64 -> base 2^26
615 mov $h0,%rax
616 mov $h0,%rdx
617 shr \$52,$h0
618 mov $h1,$d1
619 mov $h1,$d2
620 shr \$26,%rdx
621 and \$0x3ffffff,%rax # h[0]
622 shl \$12,$d1
623 and \$0x3ffffff,%rdx # h[1]
624 shr \$14,$h1
625 or $d1,$h0
626 shl \$24,$h2
627 and \$0x3ffffff,$h0 # h[2]
628 shr \$40,$d2
629 and \$0x3ffffff,$h1 # h[3]
630 or $d2,$h2 # h[4]
631
632 vmovd %rax#d,$H0
633 vmovd %rdx#d,$H1
634 vmovd $h0#d,$H2
635 vmovd $h1#d,$H3
636 vmovd $h2#d,$H4
637 movl \$1,20($ctx) # set is_base2_26
638
639 call __poly1305_init_avx
640
641 .Lproceed_avx:
642 mov %r15,$len
643
644 mov 0(%rsp),%r15
645 mov 8(%rsp),%r14
646 mov 16(%rsp),%r13
647 mov 24(%rsp),%r12
648 mov 32(%rsp),%rbp
649 mov 40(%rsp),%rbx
650 lea 48(%rsp),%rax
651 lea 48(%rsp),%rsp
652 .Lbase2_64_avx_epilogue:
653 jmp .Ldo_avx
654
655 .align 32
656 .Leven_avx:
657 vmovd 4*0($ctx),$H0 # load hash value
658 vmovd 4*1($ctx),$H1
659 vmovd 4*2($ctx),$H2
660 vmovd 4*3($ctx),$H3
661 vmovd 4*4($ctx),$H4
662
663 .Ldo_avx:
664 ___
665 $code.=<<___ if (!$win64);
666 lea -0x58(%rsp),%r11
667 sub \$0x178,%rsp
668 ___
669 $code.=<<___ if ($win64);
670 lea -0xf8(%rsp),%r11
671 sub \$0x218,%rsp
672 vmovdqa %xmm6,0x50(%r11)
673 vmovdqa %xmm7,0x60(%r11)
674 vmovdqa %xmm8,0x70(%r11)
675 vmovdqa %xmm9,0x80(%r11)
676 vmovdqa %xmm10,0x90(%r11)
677 vmovdqa %xmm11,0xa0(%r11)
678 vmovdqa %xmm12,0xb0(%r11)
679 vmovdqa %xmm13,0xc0(%r11)
680 vmovdqa %xmm14,0xd0(%r11)
681 vmovdqa %xmm15,0xe0(%r11)
682 .Ldo_avx_body:
683 ___
684 $code.=<<___;
685 sub \$64,$len
686 lea -32($inp),%rax
687 cmovc %rax,$inp
688
689 vmovdqu `16*3`($ctx),$D4 # preload r0^2
690 lea `16*3+64`($ctx),$ctx # size optimization
691 lea .Lconst(%rip),%rcx
692
693 ################################################################
694 # load input
695 vmovdqu 16*2($inp),$T0
696 vmovdqu 16*3($inp),$T1
697 vmovdqa 64(%rcx),$MASK # .Lmask26
698
699 vpsrldq \$6,$T0,$T2 # splat input
700 vpsrldq \$6,$T1,$T3
701 vpunpckhqdq $T1,$T0,$T4 # 4
702 vpunpcklqdq $T1,$T0,$T0 # 0:1
703 vpunpcklqdq $T3,$T2,$T3 # 2:3
704
705 vpsrlq \$40,$T4,$T4 # 4
706 vpsrlq \$26,$T0,$T1
707 vpand $MASK,$T0,$T0 # 0
708 vpsrlq \$4,$T3,$T2
709 vpand $MASK,$T1,$T1 # 1
710 vpsrlq \$30,$T3,$T3
711 vpand $MASK,$T2,$T2 # 2
712 vpand $MASK,$T3,$T3 # 3
713 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
714
715 jbe .Lskip_loop_avx
716
717 # expand and copy pre-calculated table to stack
718 vmovdqu `16*1-64`($ctx),$D1
719 vmovdqu `16*2-64`($ctx),$D2
720 vpshufd \$0xEE,$D4,$D3 # 34xx -> 3434
721 vpshufd \$0x44,$D4,$D0 # xx12 -> 1212
722 vmovdqa $D3,-0x90(%r11)
723 vmovdqa $D0,0x00(%rsp)
724 vpshufd \$0xEE,$D1,$D4
725 vmovdqu `16*3-64`($ctx),$D0
726 vpshufd \$0x44,$D1,$D1
727 vmovdqa $D4,-0x80(%r11)
728 vmovdqa $D1,0x10(%rsp)
729 vpshufd \$0xEE,$D2,$D3
730 vmovdqu `16*4-64`($ctx),$D1
731 vpshufd \$0x44,$D2,$D2
732 vmovdqa $D3,-0x70(%r11)
733 vmovdqa $D2,0x20(%rsp)
734 vpshufd \$0xEE,$D0,$D4
735 vmovdqu `16*5-64`($ctx),$D2
736 vpshufd \$0x44,$D0,$D0
737 vmovdqa $D4,-0x60(%r11)
738 vmovdqa $D0,0x30(%rsp)
739 vpshufd \$0xEE,$D1,$D3
740 vmovdqu `16*6-64`($ctx),$D0
741 vpshufd \$0x44,$D1,$D1
742 vmovdqa $D3,-0x50(%r11)
743 vmovdqa $D1,0x40(%rsp)
744 vpshufd \$0xEE,$D2,$D4
745 vmovdqu `16*7-64`($ctx),$D1
746 vpshufd \$0x44,$D2,$D2
747 vmovdqa $D4,-0x40(%r11)
748 vmovdqa $D2,0x50(%rsp)
749 vpshufd \$0xEE,$D0,$D3
750 vmovdqu `16*8-64`($ctx),$D2
751 vpshufd \$0x44,$D0,$D0
752 vmovdqa $D3,-0x30(%r11)
753 vmovdqa $D0,0x60(%rsp)
754 vpshufd \$0xEE,$D1,$D4
755 vpshufd \$0x44,$D1,$D1
756 vmovdqa $D4,-0x20(%r11)
757 vmovdqa $D1,0x70(%rsp)
758 vpshufd \$0xEE,$D2,$D3
759 vmovdqa 0x00(%rsp),$D4 # preload r0^2
760 vpshufd \$0x44,$D2,$D2
761 vmovdqa $D3,-0x10(%r11)
762 vmovdqa $D2,0x80(%rsp)
763
764 jmp .Loop_avx
765
766 .align 32
767 .Loop_avx:
768 ################################################################
769 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
770 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
771 # \___________________/
772 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
773 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
774 # \___________________/ \____________________/
775 #
776 # Note that we start with inp[2:3]*r^2. This is because it
777 # doesn't depend on reduction in previous iteration.
778 ################################################################
779 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
780 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
781 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
782 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
783 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
784 #
785 # though note that $Tx and $Hx are "reversed" in this section,
786 # and $D4 is preloaded with r0^2...
787
788 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
789 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
790 vmovdqa $H2,0x20(%r11) # offload hash
791 vpmuludq $T2,$D4,$D2 # d3 = h2*r0
792 vmovdqa 0x10(%rsp),$H2 # r1^2
793 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
794 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
795
796 vmovdqa $H0,0x00(%r11) #
797 vpmuludq 0x20(%rsp),$T4,$H0 # h4*s1
798 vmovdqa $H1,0x10(%r11) #
799 vpmuludq $T3,$H2,$H1 # h3*r1
800 vpaddq $H0,$D0,$D0 # d0 += h4*s1
801 vpaddq $H1,$D4,$D4 # d4 += h3*r1
802 vmovdqa $H3,0x30(%r11) #
803 vpmuludq $T2,$H2,$H0 # h2*r1
804 vpmuludq $T1,$H2,$H1 # h1*r1
805 vpaddq $H0,$D3,$D3 # d3 += h2*r1
806 vmovdqa 0x30(%rsp),$H3 # r2^2
807 vpaddq $H1,$D2,$D2 # d2 += h1*r1
808 vmovdqa $H4,0x40(%r11) #
809 vpmuludq $T0,$H2,$H2 # h0*r1
810 vpmuludq $T2,$H3,$H0 # h2*r2
811 vpaddq $H2,$D1,$D1 # d1 += h0*r1
812
813 vmovdqa 0x40(%rsp),$H4 # s2^2
814 vpaddq $H0,$D4,$D4 # d4 += h2*r2
815 vpmuludq $T1,$H3,$H1 # h1*r2
816 vpmuludq $T0,$H3,$H3 # h0*r2
817 vpaddq $H1,$D3,$D3 # d3 += h1*r2
818 vmovdqa 0x50(%rsp),$H2 # r3^2
819 vpaddq $H3,$D2,$D2 # d2 += h0*r2
820 vpmuludq $T4,$H4,$H0 # h4*s2
821 vpmuludq $T3,$H4,$H4 # h3*s2
822 vpaddq $H0,$D1,$D1 # d1 += h4*s2
823 vmovdqa 0x60(%rsp),$H3 # s3^2
824 vpaddq $H4,$D0,$D0 # d0 += h3*s2
825
826 vmovdqa 0x80(%rsp),$H4 # s4^2
827 vpmuludq $T1,$H2,$H1 # h1*r3
828 vpmuludq $T0,$H2,$H2 # h0*r3
829 vpaddq $H1,$D4,$D4 # d4 += h1*r3
830 vpaddq $H2,$D3,$D3 # d3 += h0*r3
831 vpmuludq $T4,$H3,$H0 # h4*s3
832 vpmuludq $T3,$H3,$H1 # h3*s3
833 vpaddq $H0,$D2,$D2 # d2 += h4*s3
834 vmovdqu 16*0($inp),$H0 # load input
835 vpaddq $H1,$D1,$D1 # d1 += h3*s3
836 vpmuludq $T2,$H3,$H3 # h2*s3
837 vpmuludq $T2,$H4,$T2 # h2*s4
838 vpaddq $H3,$D0,$D0 # d0 += h2*s3
839
840 vmovdqu 16*1($inp),$H1 #
841 vpaddq $T2,$D1,$D1 # d1 += h2*s4
842 vpmuludq $T3,$H4,$T3 # h3*s4
843 vpmuludq $T4,$H4,$T4 # h4*s4
844 vpsrldq \$6,$H0,$H2 # splat input
845 vpaddq $T3,$D2,$D2 # d2 += h3*s4
846 vpaddq $T4,$D3,$D3 # d3 += h4*s4
847 vpsrldq \$6,$H1,$H3 #
848 vpmuludq 0x70(%rsp),$T0,$T4 # h0*r4
849 vpmuludq $T1,$H4,$T0 # h1*s4
850 vpunpckhqdq $H1,$H0,$H4 # 4
851 vpaddq $T4,$D4,$D4 # d4 += h0*r4
852 vmovdqa -0x90(%r11),$T4 # r0^4
853 vpaddq $T0,$D0,$D0 # d0 += h1*s4
854
855 vpunpcklqdq $H1,$H0,$H0 # 0:1
856 vpunpcklqdq $H3,$H2,$H3 # 2:3
857
858 #vpsrlq \$40,$H4,$H4 # 4
859 vpsrldq \$`40/8`,$H4,$H4 # 4
860 vpsrlq \$26,$H0,$H1
861 vpand $MASK,$H0,$H0 # 0
862 vpsrlq \$4,$H3,$H2
863 vpand $MASK,$H1,$H1 # 1
864 vpand 0(%rcx),$H4,$H4 # .Lmask24
865 vpsrlq \$30,$H3,$H3
866 vpand $MASK,$H2,$H2 # 2
867 vpand $MASK,$H3,$H3 # 3
868 vpor 32(%rcx),$H4,$H4 # padbit, yes, always
869
870 vpaddq 0x00(%r11),$H0,$H0 # add hash value
871 vpaddq 0x10(%r11),$H1,$H1
872 vpaddq 0x20(%r11),$H2,$H2
873 vpaddq 0x30(%r11),$H3,$H3
874 vpaddq 0x40(%r11),$H4,$H4
875
876 lea 16*2($inp),%rax
877 lea 16*4($inp),$inp
878 sub \$64,$len
879 cmovc %rax,$inp
880
881 ################################################################
882 # Now we accumulate (inp[0:1]+hash)*r^4
883 ################################################################
884 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
885 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
886 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
887 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
888 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
889
890 vpmuludq $H0,$T4,$T0 # h0*r0
891 vpmuludq $H1,$T4,$T1 # h1*r0
892 vpaddq $T0,$D0,$D0
893 vpaddq $T1,$D1,$D1
894 vmovdqa -0x80(%r11),$T2 # r1^4
895 vpmuludq $H2,$T4,$T0 # h2*r0
896 vpmuludq $H3,$T4,$T1 # h3*r0
897 vpaddq $T0,$D2,$D2
898 vpaddq $T1,$D3,$D3
899 vpmuludq $H4,$T4,$T4 # h4*r0
900 vpmuludq -0x70(%r11),$H4,$T0 # h4*s1
901 vpaddq $T4,$D4,$D4
902
903 vpaddq $T0,$D0,$D0 # d0 += h4*s1
904 vpmuludq $H2,$T2,$T1 # h2*r1
905 vpmuludq $H3,$T2,$T0 # h3*r1
906 vpaddq $T1,$D3,$D3 # d3 += h2*r1
907 vmovdqa -0x60(%r11),$T3 # r2^4
908 vpaddq $T0,$D4,$D4 # d4 += h3*r1
909 vpmuludq $H1,$T2,$T1 # h1*r1
910 vpmuludq $H0,$T2,$T2 # h0*r1
911 vpaddq $T1,$D2,$D2 # d2 += h1*r1
912 vpaddq $T2,$D1,$D1 # d1 += h0*r1
913
914 vmovdqa -0x50(%r11),$T4 # s2^4
915 vpmuludq $H2,$T3,$T0 # h2*r2
916 vpmuludq $H1,$T3,$T1 # h1*r2
917 vpaddq $T0,$D4,$D4 # d4 += h2*r2
918 vpaddq $T1,$D3,$D3 # d3 += h1*r2
919 vmovdqa -0x40(%r11),$T2 # r3^4
920 vpmuludq $H0,$T3,$T3 # h0*r2
921 vpmuludq $H4,$T4,$T0 # h4*s2
922 vpaddq $T3,$D2,$D2 # d2 += h0*r2
923 vpaddq $T0,$D1,$D1 # d1 += h4*s2
924 vmovdqa -0x30(%r11),$T3 # s3^4
925 vpmuludq $H3,$T4,$T4 # h3*s2
926 vpmuludq $H1,$T2,$T1 # h1*r3
927 vpaddq $T4,$D0,$D0 # d0 += h3*s2
928
929 vmovdqa -0x10(%r11),$T4 # s4^4
930 vpaddq $T1,$D4,$D4 # d4 += h1*r3
931 vpmuludq $H0,$T2,$T2 # h0*r3
932 vpmuludq $H4,$T3,$T0 # h4*s3
933 vpaddq $T2,$D3,$D3 # d3 += h0*r3
934 vpaddq $T0,$D2,$D2 # d2 += h4*s3
935 vmovdqu 16*2($inp),$T0 # load input
936 vpmuludq $H3,$T3,$T2 # h3*s3
937 vpmuludq $H2,$T3,$T3 # h2*s3
938 vpaddq $T2,$D1,$D1 # d1 += h3*s3
939 vmovdqu 16*3($inp),$T1 #
940 vpaddq $T3,$D0,$D0 # d0 += h2*s3
941
942 vpmuludq $H2,$T4,$H2 # h2*s4
943 vpmuludq $H3,$T4,$H3 # h3*s4
944 vpsrldq \$6,$T0,$T2 # splat input
945 vpaddq $H2,$D1,$D1 # d1 += h2*s4
946 vpmuludq $H4,$T4,$H4 # h4*s4
947 vpsrldq \$6,$T1,$T3 #
948 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*s4
949 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*s4
950 vpmuludq -0x20(%r11),$H0,$H4 # h0*r4
951 vpmuludq $H1,$T4,$H0
952 vpunpckhqdq $T1,$T0,$T4 # 4
953 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
954 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
955
956 vpunpcklqdq $T1,$T0,$T0 # 0:1
957 vpunpcklqdq $T3,$T2,$T3 # 2:3
958
959 #vpsrlq \$40,$T4,$T4 # 4
960 vpsrldq \$`40/8`,$T4,$T4 # 4
961 vpsrlq \$26,$T0,$T1
962 vmovdqa 0x00(%rsp),$D4 # preload r0^2
963 vpand $MASK,$T0,$T0 # 0
964 vpsrlq \$4,$T3,$T2
965 vpand $MASK,$T1,$T1 # 1
966 vpand 0(%rcx),$T4,$T4 # .Lmask24
967 vpsrlq \$30,$T3,$T3
968 vpand $MASK,$T2,$T2 # 2
969 vpand $MASK,$T3,$T3 # 3
970 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
971
972 ################################################################
973 # lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
974 # and P. Schwabe
975
976 vpsrlq \$26,$H3,$D3
977 vpand $MASK,$H3,$H3
978 vpaddq $D3,$H4,$H4 # h3 -> h4
979
980 vpsrlq \$26,$H0,$D0
981 vpand $MASK,$H0,$H0
982 vpaddq $D0,$D1,$H1 # h0 -> h1
983
984 vpsrlq \$26,$H4,$D0
985 vpand $MASK,$H4,$H4
986
987 vpsrlq \$26,$H1,$D1
988 vpand $MASK,$H1,$H1
989 vpaddq $D1,$H2,$H2 # h1 -> h2
990
991 vpaddq $D0,$H0,$H0
992 vpsllq \$2,$D0,$D0
993 vpaddq $D0,$H0,$H0 # h4 -> h0
994
995 vpsrlq \$26,$H2,$D2
996 vpand $MASK,$H2,$H2
997 vpaddq $D2,$H3,$H3 # h2 -> h3
998
999 vpsrlq \$26,$H0,$D0
1000 vpand $MASK,$H0,$H0
1001 vpaddq $D0,$H1,$H1 # h0 -> h1
1002
1003 vpsrlq \$26,$H3,$D3
1004 vpand $MASK,$H3,$H3
1005 vpaddq $D3,$H4,$H4 # h3 -> h4
1006
1007 ja .Loop_avx
1008
1009 .Lskip_loop_avx:
1010 ################################################################
1011 # multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
1012
1013 vpshufd \$0x10,$D4,$D4 # r0^n, xx12 -> x1x2
1014 add \$32,$len
1015 jnz .Long_tail_avx
1016
1017 vpaddq $H2,$T2,$T2
1018 vpaddq $H0,$T0,$T0
1019 vpaddq $H1,$T1,$T1
1020 vpaddq $H3,$T3,$T3
1021 vpaddq $H4,$T4,$T4
1022
1023 .Long_tail_avx:
1024 vmovdqa $H2,0x20(%r11)
1025 vmovdqa $H0,0x00(%r11)
1026 vmovdqa $H1,0x10(%r11)
1027 vmovdqa $H3,0x30(%r11)
1028 vmovdqa $H4,0x40(%r11)
1029
1030 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1031 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1032 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1033 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1034 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1035
1036 vpmuludq $T2,$D4,$D2 # d2 = h2*r0
1037 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
1038 vpshufd \$0x10,`16*1-64`($ctx),$H2 # r1^n
1039 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
1040 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
1041 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
1042
1043 vpmuludq $T3,$H2,$H0 # h3*r1
1044 vpaddq $H0,$D4,$D4 # d4 += h3*r1
1045 vpshufd \$0x10,`16*2-64`($ctx),$H3 # s1^n
1046 vpmuludq $T2,$H2,$H1 # h2*r1
1047 vpaddq $H1,$D3,$D3 # d3 += h2*r1
1048 vpshufd \$0x10,`16*3-64`($ctx),$H4 # r2^n
1049 vpmuludq $T1,$H2,$H0 # h1*r1
1050 vpaddq $H0,$D2,$D2 # d2 += h1*r1
1051 vpmuludq $T0,$H2,$H2 # h0*r1
1052 vpaddq $H2,$D1,$D1 # d1 += h0*r1
1053 vpmuludq $T4,$H3,$H3 # h4*s1
1054 vpaddq $H3,$D0,$D0 # d0 += h4*s1
1055
1056 vpshufd \$0x10,`16*4-64`($ctx),$H2 # s2^n
1057 vpmuludq $T2,$H4,$H1 # h2*r2
1058 vpaddq $H1,$D4,$D4 # d4 += h2*r2
1059 vpmuludq $T1,$H4,$H0 # h1*r2
1060 vpaddq $H0,$D3,$D3 # d3 += h1*r2
1061 vpshufd \$0x10,`16*5-64`($ctx),$H3 # r3^n
1062 vpmuludq $T0,$H4,$H4 # h0*r2
1063 vpaddq $H4,$D2,$D2 # d2 += h0*r2
1064 vpmuludq $T4,$H2,$H1 # h4*s2
1065 vpaddq $H1,$D1,$D1 # d1 += h4*s2
1066 vpshufd \$0x10,`16*6-64`($ctx),$H4 # s3^n
1067 vpmuludq $T3,$H2,$H2 # h3*s2
1068 vpaddq $H2,$D0,$D0 # d0 += h3*s2
1069
1070 vpmuludq $T1,$H3,$H0 # h1*r3
1071 vpaddq $H0,$D4,$D4 # d4 += h1*r3
1072 vpmuludq $T0,$H3,$H3 # h0*r3
1073 vpaddq $H3,$D3,$D3 # d3 += h0*r3
1074 vpshufd \$0x10,`16*7-64`($ctx),$H2 # r4^n
1075 vpmuludq $T4,$H4,$H1 # h4*s3
1076 vpaddq $H1,$D2,$D2 # d2 += h4*s3
1077 vpshufd \$0x10,`16*8-64`($ctx),$H3 # s4^n
1078 vpmuludq $T3,$H4,$H0 # h3*s3
1079 vpaddq $H0,$D1,$D1 # d1 += h3*s3
1080 vpmuludq $T2,$H4,$H4 # h2*s3
1081 vpaddq $H4,$D0,$D0 # d0 += h2*s3
1082
1083 vpmuludq $T0,$H2,$H2 # h0*r4
1084 vpaddq $H2,$D4,$D4 # h4 = d4 + h0*r4
1085 vpmuludq $T4,$H3,$H1 # h4*s4
1086 vpaddq $H1,$D3,$D3 # h3 = d3 + h4*s4
1087 vpmuludq $T3,$H3,$H0 # h3*s4
1088 vpaddq $H0,$D2,$D2 # h2 = d2 + h3*s4
1089 vpmuludq $T2,$H3,$H1 # h2*s4
1090 vpaddq $H1,$D1,$D1 # h1 = d1 + h2*s4
1091 vpmuludq $T1,$H3,$H3 # h1*s4
1092 vpaddq $H3,$D0,$D0 # h0 = d0 + h1*s4
1093
1094 jz .Lshort_tail_avx
1095
1096 vmovdqu 16*0($inp),$H0 # load input
1097 vmovdqu 16*1($inp),$H1
1098
1099 vpsrldq \$6,$H0,$H2 # splat input
1100 vpsrldq \$6,$H1,$H3
1101 vpunpckhqdq $H1,$H0,$H4 # 4
1102 vpunpcklqdq $H1,$H0,$H0 # 0:1
1103 vpunpcklqdq $H3,$H2,$H3 # 2:3
1104
1105 vpsrlq \$40,$H4,$H4 # 4
1106 vpsrlq \$26,$H0,$H1
1107 vpand $MASK,$H0,$H0 # 0
1108 vpsrlq \$4,$H3,$H2
1109 vpand $MASK,$H1,$H1 # 1
1110 vpsrlq \$30,$H3,$H3
1111 vpand $MASK,$H2,$H2 # 2
1112 vpand $MASK,$H3,$H3 # 3
1113 vpor 32(%rcx),$H4,$H4 # padbit, yes, always
1114
1115 vpshufd \$0x32,`16*0-64`($ctx),$T4 # r0^n, 34xx -> x3x4
1116 vpaddq 0x00(%r11),$H0,$H0
1117 vpaddq 0x10(%r11),$H1,$H1
1118 vpaddq 0x20(%r11),$H2,$H2
1119 vpaddq 0x30(%r11),$H3,$H3
1120 vpaddq 0x40(%r11),$H4,$H4
1121
1122 ################################################################
1123 # multiply (inp[0:1]+hash) by r^4:r^3 and accumulate
1124
1125 vpmuludq $H0,$T4,$T0 # h0*r0
1126 vpaddq $T0,$D0,$D0 # d0 += h0*r0
1127 vpmuludq $H1,$T4,$T1 # h1*r0
1128 vpaddq $T1,$D1,$D1 # d1 += h1*r0
1129 vpmuludq $H2,$T4,$T0 # h2*r0
1130 vpaddq $T0,$D2,$D2 # d2 += h2*r0
1131 vpshufd \$0x32,`16*1-64`($ctx),$T2 # r1^n
1132 vpmuludq $H3,$T4,$T1 # h3*r0
1133 vpaddq $T1,$D3,$D3 # d3 += h3*r0
1134 vpmuludq $H4,$T4,$T4 # h4*r0
1135 vpaddq $T4,$D4,$D4 # d4 += h4*r0
1136
1137 vpmuludq $H3,$T2,$T0 # h3*r1
1138 vpaddq $T0,$D4,$D4 # d4 += h3*r1
1139 vpshufd \$0x32,`16*2-64`($ctx),$T3 # s1
1140 vpmuludq $H2,$T2,$T1 # h2*r1
1141 vpaddq $T1,$D3,$D3 # d3 += h2*r1
1142 vpshufd \$0x32,`16*3-64`($ctx),$T4 # r2
1143 vpmuludq $H1,$T2,$T0 # h1*r1
1144 vpaddq $T0,$D2,$D2 # d2 += h1*r1
1145 vpmuludq $H0,$T2,$T2 # h0*r1
1146 vpaddq $T2,$D1,$D1 # d1 += h0*r1
1147 vpmuludq $H4,$T3,$T3 # h4*s1
1148 vpaddq $T3,$D0,$D0 # d0 += h4*s1
1149
1150 vpshufd \$0x32,`16*4-64`($ctx),$T2 # s2
1151 vpmuludq $H2,$T4,$T1 # h2*r2
1152 vpaddq $T1,$D4,$D4 # d4 += h2*r2
1153 vpmuludq $H1,$T4,$T0 # h1*r2
1154 vpaddq $T0,$D3,$D3 # d3 += h1*r2
1155 vpshufd \$0x32,`16*5-64`($ctx),$T3 # r3
1156 vpmuludq $H0,$T4,$T4 # h0*r2
1157 vpaddq $T4,$D2,$D2 # d2 += h0*r2
1158 vpmuludq $H4,$T2,$T1 # h4*s2
1159 vpaddq $T1,$D1,$D1 # d1 += h4*s2
1160 vpshufd \$0x32,`16*6-64`($ctx),$T4 # s3
1161 vpmuludq $H3,$T2,$T2 # h3*s2
1162 vpaddq $T2,$D0,$D0 # d0 += h3*s2
1163
1164 vpmuludq $H1,$T3,$T0 # h1*r3
1165 vpaddq $T0,$D4,$D4 # d4 += h1*r3
1166 vpmuludq $H0,$T3,$T3 # h0*r3
1167 vpaddq $T3,$D3,$D3 # d3 += h0*r3
1168 vpshufd \$0x32,`16*7-64`($ctx),$T2 # r4
1169 vpmuludq $H4,$T4,$T1 # h4*s3
1170 vpaddq $T1,$D2,$D2 # d2 += h4*s3
1171 vpshufd \$0x32,`16*8-64`($ctx),$T3 # s4
1172 vpmuludq $H3,$T4,$T0 # h3*s3
1173 vpaddq $T0,$D1,$D1 # d1 += h3*s3
1174 vpmuludq $H2,$T4,$T4 # h2*s3
1175 vpaddq $T4,$D0,$D0 # d0 += h2*s3
1176
1177 vpmuludq $H0,$T2,$T2 # h0*r4
1178 vpaddq $T2,$D4,$D4 # d4 += h0*r4
1179 vpmuludq $H4,$T3,$T1 # h4*s4
1180 vpaddq $T1,$D3,$D3 # d3 += h4*s4
1181 vpmuludq $H3,$T3,$T0 # h3*s4
1182 vpaddq $T0,$D2,$D2 # d2 += h3*s4
1183 vpmuludq $H2,$T3,$T1 # h2*s4
1184 vpaddq $T1,$D1,$D1 # d1 += h2*s4
1185 vpmuludq $H1,$T3,$T3 # h1*s4
1186 vpaddq $T3,$D0,$D0 # d0 += h1*s4
1187
1188 .Lshort_tail_avx:
1189 ################################################################
1190 # horizontal addition
1191
1192 vpsrldq \$8,$D4,$T4
1193 vpsrldq \$8,$D3,$T3
1194 vpsrldq \$8,$D1,$T1
1195 vpsrldq \$8,$D0,$T0
1196 vpsrldq \$8,$D2,$T2
1197 vpaddq $T3,$D3,$D3
1198 vpaddq $T4,$D4,$D4
1199 vpaddq $T0,$D0,$D0
1200 vpaddq $T1,$D1,$D1
1201 vpaddq $T2,$D2,$D2
1202
1203 ################################################################
1204 # lazy reduction
1205
1206 vpsrlq \$26,$D3,$H3
1207 vpand $MASK,$D3,$D3
1208 vpaddq $H3,$D4,$D4 # h3 -> h4
1209
1210 vpsrlq \$26,$D0,$H0
1211 vpand $MASK,$D0,$D0
1212 vpaddq $H0,$D1,$D1 # h0 -> h1
1213
1214 vpsrlq \$26,$D4,$H4
1215 vpand $MASK,$D4,$D4
1216
1217 vpsrlq \$26,$D1,$H1
1218 vpand $MASK,$D1,$D1
1219 vpaddq $H1,$D2,$D2 # h1 -> h2
1220
1221 vpaddq $H4,$D0,$D0
1222 vpsllq \$2,$H4,$H4
1223 vpaddq $H4,$D0,$D0 # h4 -> h0
1224
1225 vpsrlq \$26,$D2,$H2
1226 vpand $MASK,$D2,$D2
1227 vpaddq $H2,$D3,$D3 # h2 -> h3
1228
1229 vpsrlq \$26,$D0,$H0
1230 vpand $MASK,$D0,$D0
1231 vpaddq $H0,$D1,$D1 # h0 -> h1
1232
1233 vpsrlq \$26,$D3,$H3
1234 vpand $MASK,$D3,$D3
1235 vpaddq $H3,$D4,$D4 # h3 -> h4
1236
1237 vmovd $D0,`4*0-48-64`($ctx) # save partially reduced
1238 vmovd $D1,`4*1-48-64`($ctx)
1239 vmovd $D2,`4*2-48-64`($ctx)
1240 vmovd $D3,`4*3-48-64`($ctx)
1241 vmovd $D4,`4*4-48-64`($ctx)
1242 ___
1243 $code.=<<___ if ($win64);
1244 vmovdqa 0x50(%r11),%xmm6
1245 vmovdqa 0x60(%r11),%xmm7
1246 vmovdqa 0x70(%r11),%xmm8
1247 vmovdqa 0x80(%r11),%xmm9
1248 vmovdqa 0x90(%r11),%xmm10
1249 vmovdqa 0xa0(%r11),%xmm11
1250 vmovdqa 0xb0(%r11),%xmm12
1251 vmovdqa 0xc0(%r11),%xmm13
1252 vmovdqa 0xd0(%r11),%xmm14
1253 vmovdqa 0xe0(%r11),%xmm15
1254 lea 0xf8(%r11),%rsp
1255 .Ldo_avx_epilogue:
1256 ___
1257 $code.=<<___ if (!$win64);
1258 lea 0x58(%r11),%rsp
1259 ___
1260 $code.=<<___;
1261 vzeroupper
1262 ret
1263 .size poly1305_blocks_avx,.-poly1305_blocks_avx
1264
1265 .type poly1305_emit_avx,\@function,3
1266 .align 32
1267 poly1305_emit_avx:
1268 cmpl \$0,20($ctx) # is_base2_26?
1269 je .Lemit
1270
1271 mov 0($ctx),%eax # load hash value base 2^26
1272 mov 4($ctx),%ecx
1273 mov 8($ctx),%r8d
1274 mov 12($ctx),%r11d
1275 mov 16($ctx),%r10d
1276
1277 shl \$26,%rcx # base 2^26 -> base 2^64
1278 mov %r8,%r9
1279 shl \$52,%r8
1280 add %rcx,%rax
1281 shr \$12,%r9
1282 add %rax,%r8 # h0
1283 adc \$0,%r9
1284
1285 shl \$14,%r11
1286 mov %r10,%rax
1287 shr \$24,%r10
1288 add %r11,%r9
1289 shl \$40,%rax
1290 add %rax,%r9 # h1
1291 adc \$0,%r10 # h2
1292
1293 mov %r10,%rax # could be partially reduced, so reduce
1294 mov %r10,%rcx
1295 and \$3,%r10
1296 shr \$2,%rax
1297 and \$-4,%rcx
1298 add %rcx,%rax
1299 add %rax,%r8
1300 adc \$0,%r9
1301
1302 mov %r8,%rax
1303 add \$5,%r8 # compare to modulus
1304 mov %r9,%rcx
1305 adc \$0,%r9
1306 adc \$0,%r10
1307 shr \$2,%r10 # did 130-bit value overfow?
1308 cmovnz %r8,%rax
1309 cmovnz %r9,%rcx
1310
1311 add 0($nonce),%rax # accumulate nonce
1312 adc 8($nonce),%rcx
1313 mov %rax,0($mac) # write result
1314 mov %rcx,8($mac)
1315
1316 ret
1317 .size poly1305_emit_avx,.-poly1305_emit_avx
1318 ___
1319
1320 if ($avx>1) {
1321 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1322 map("%ymm$_",(0..15));
1323 my $S4=$MASK;
1324
1325 $code.=<<___;
1326 .type poly1305_blocks_avx2,\@function,4
1327 .align 32
1328 poly1305_blocks_avx2:
1329 mov 20($ctx),%r8d # is_base2_26
1330 cmp \$128,$len
1331 jae .Lblocks_avx2
1332 test %r8d,%r8d
1333 jz .Lblocks
1334
1335 .Lblocks_avx2:
1336 and \$-16,$len
1337 jz .Lno_data_avx2
1338
1339 vzeroupper
1340
1341 test %r8d,%r8d
1342 jz .Lbase2_64_avx2
1343
1344 test \$63,$len
1345 jz .Leven_avx2
1346
1347 push %rbx
1348 push %rbp
1349 push %r12
1350 push %r13
1351 push %r14
1352 push %r15
1353 .Lblocks_avx2_body:
1354
1355 mov $len,%r15 # reassign $len
1356
1357 mov 0($ctx),$d1 # load hash value
1358 mov 8($ctx),$d2
1359 mov 16($ctx),$h2#d
1360
1361 mov 24($ctx),$r0 # load r
1362 mov 32($ctx),$s1
1363
1364 ################################# base 2^26 -> base 2^64
1365 mov $d1#d,$h0#d
1366 and \$-1<<31,$d1
1367 mov $d2,$r1 # borrow $r1
1368 mov $d2#d,$h1#d
1369 and \$-1<<31,$d2
1370
1371 shr \$6,$d1
1372 shl \$52,$r1
1373 add $d1,$h0
1374 shr \$12,$h1
1375 shr \$18,$d2
1376 add $r1,$h0
1377 adc $d2,$h1
1378
1379 mov $h2,$d1
1380 shl \$40,$d1
1381 shr \$24,$h2
1382 add $d1,$h1
1383 adc \$0,$h2 # can be partially reduced...
1384
1385 mov \$-4,$d2 # ... so reduce
1386 mov $h2,$d1
1387 and $h2,$d2
1388 shr \$2,$d1
1389 and \$3,$h2
1390 add $d2,$d1 # =*5
1391 add $d1,$h0
1392 adc \$0,$h1
1393
1394 mov $s1,$r1
1395 mov $s1,%rax
1396 shr \$2,$s1
1397 add $r1,$s1 # s1 = r1 + (r1 >> 2)
1398
1399 .Lbase2_26_pre_avx2:
1400 add 0($inp),$h0 # accumulate input
1401 adc 8($inp),$h1
1402 lea 16($inp),$inp
1403 adc $padbit,$h2
1404 sub \$16,%r15
1405
1406 call __poly1305_block
1407 mov $r1,%rax
1408
1409 test \$63,%r15
1410 jnz .Lbase2_26_pre_avx2
1411
1412 test $padbit,$padbit # if $padbit is zero,
1413 jz .Lstore_base2_64_avx2 # store hash in base 2^64 format
1414
1415 ################################# base 2^64 -> base 2^26
1416 mov $h0,%rax
1417 mov $h0,%rdx
1418 shr \$52,$h0
1419 mov $h1,$r0
1420 mov $h1,$r1
1421 shr \$26,%rdx
1422 and \$0x3ffffff,%rax # h[0]
1423 shl \$12,$r0
1424 and \$0x3ffffff,%rdx # h[1]
1425 shr \$14,$h1
1426 or $r0,$h0
1427 shl \$24,$h2
1428 and \$0x3ffffff,$h0 # h[2]
1429 shr \$40,$r1
1430 and \$0x3ffffff,$h1 # h[3]
1431 or $r1,$h2 # h[4]
1432
1433 test %r15,%r15
1434 jz .Lstore_base2_26_avx2
1435
1436 vmovd %rax#d,%x#$H0
1437 vmovd %rdx#d,%x#$H1
1438 vmovd $h0#d,%x#$H2
1439 vmovd $h1#d,%x#$H3
1440 vmovd $h2#d,%x#$H4
1441 jmp .Lproceed_avx2
1442
1443 .align 32
1444 .Lstore_base2_64_avx2:
1445 mov $h0,0($ctx)
1446 mov $h1,8($ctx)
1447 mov $h2,16($ctx) # note that is_base2_26 is zeroed
1448 jmp .Ldone_avx2
1449
1450 .align 16
1451 .Lstore_base2_26_avx2:
1452 mov %rax#d,0($ctx) # store hash value base 2^26
1453 mov %rdx#d,4($ctx)
1454 mov $h0#d,8($ctx)
1455 mov $h1#d,12($ctx)
1456 mov $h2#d,16($ctx)
1457 .align 16
1458 .Ldone_avx2:
1459 mov 0(%rsp),%r15
1460 mov 8(%rsp),%r14
1461 mov 16(%rsp),%r13
1462 mov 24(%rsp),%r12
1463 mov 32(%rsp),%rbp
1464 mov 40(%rsp),%rbx
1465 lea 48(%rsp),%rsp
1466 .Lno_data_avx2:
1467 .Lblocks_avx2_epilogue:
1468 ret
1469
1470 .align 32
1471 .Lbase2_64_avx2:
1472 push %rbx
1473 push %rbp
1474 push %r12
1475 push %r13
1476 push %r14
1477 push %r15
1478 .Lbase2_64_avx2_body:
1479
1480 mov $len,%r15 # reassign $len
1481
1482 mov 24($ctx),$r0 # load r
1483 mov 32($ctx),$s1
1484
1485 mov 0($ctx),$h0 # load hash value
1486 mov 8($ctx),$h1
1487 mov 16($ctx),$h2#d
1488
1489 mov $s1,$r1
1490 mov $s1,%rax
1491 shr \$2,$s1
1492 add $r1,$s1 # s1 = r1 + (r1 >> 2)
1493
1494 test \$63,$len
1495 jz .Linit_avx2
1496
1497 .Lbase2_64_pre_avx2:
1498 add 0($inp),$h0 # accumulate input
1499 adc 8($inp),$h1
1500 lea 16($inp),$inp
1501 adc $padbit,$h2
1502 sub \$16,%r15
1503
1504 call __poly1305_block
1505 mov $r1,%rax
1506
1507 test \$63,%r15
1508 jnz .Lbase2_64_pre_avx2
1509
1510 .Linit_avx2:
1511 ################################# base 2^64 -> base 2^26
1512 mov $h0,%rax
1513 mov $h0,%rdx
1514 shr \$52,$h0
1515 mov $h1,$d1
1516 mov $h1,$d2
1517 shr \$26,%rdx
1518 and \$0x3ffffff,%rax # h[0]
1519 shl \$12,$d1
1520 and \$0x3ffffff,%rdx # h[1]
1521 shr \$14,$h1
1522 or $d1,$h0
1523 shl \$24,$h2
1524 and \$0x3ffffff,$h0 # h[2]
1525 shr \$40,$d2
1526 and \$0x3ffffff,$h1 # h[3]
1527 or $d2,$h2 # h[4]
1528
1529 vmovd %rax#d,%x#$H0
1530 vmovd %rdx#d,%x#$H1
1531 vmovd $h0#d,%x#$H2
1532 vmovd $h1#d,%x#$H3
1533 vmovd $h2#d,%x#$H4
1534 movl \$1,20($ctx) # set is_base2_26
1535
1536 call __poly1305_init_avx
1537
1538 .Lproceed_avx2:
1539 mov %r15,$len
1540
1541 mov 0(%rsp),%r15
1542 mov 8(%rsp),%r14
1543 mov 16(%rsp),%r13
1544 mov 24(%rsp),%r12
1545 mov 32(%rsp),%rbp
1546 mov 40(%rsp),%rbx
1547 lea 48(%rsp),%rax
1548 lea 48(%rsp),%rsp
1549 .Lbase2_64_avx2_epilogue:
1550 jmp .Ldo_avx2
1551
1552 .align 32
1553 .Leven_avx2:
1554 vmovd 4*0($ctx),%x#$H0 # load hash value base 2^26
1555 vmovd 4*1($ctx),%x#$H1
1556 vmovd 4*2($ctx),%x#$H2
1557 vmovd 4*3($ctx),%x#$H3
1558 vmovd 4*4($ctx),%x#$H4
1559
1560 .Ldo_avx2:
1561 ___
1562 $code.=<<___ if (!$win64);
1563 lea -8(%rsp),%r11
1564 sub \$0x128,%rsp
1565 ___
1566 $code.=<<___ if ($win64);
1567 lea -0xf8(%rsp),%r11
1568 sub \$0x1c8,%rsp
1569 vmovdqa %xmm6,0x50(%r11)
1570 vmovdqa %xmm7,0x60(%r11)
1571 vmovdqa %xmm8,0x70(%r11)
1572 vmovdqa %xmm9,0x80(%r11)
1573 vmovdqa %xmm10,0x90(%r11)
1574 vmovdqa %xmm11,0xa0(%r11)
1575 vmovdqa %xmm12,0xb0(%r11)
1576 vmovdqa %xmm13,0xc0(%r11)
1577 vmovdqa %xmm14,0xd0(%r11)
1578 vmovdqa %xmm15,0xe0(%r11)
1579 .Ldo_avx2_body:
1580 ___
1581 $code.=<<___;
1582 lea 48+64($ctx),$ctx # size optimization
1583 lea .Lconst(%rip),%rcx
1584
1585 # expand and copy pre-calculated table to stack
1586 vmovdqu `16*0-64`($ctx),%x#$T2
1587 and \$-512,%rsp
1588 vmovdqu `16*1-64`($ctx),%x#$T3
1589 vmovdqu `16*2-64`($ctx),%x#$T4
1590 vmovdqu `16*3-64`($ctx),%x#$D0
1591 vmovdqu `16*4-64`($ctx),%x#$D1
1592 vmovdqu `16*5-64`($ctx),%x#$D2
1593 vmovdqu `16*6-64`($ctx),%x#$D3
1594 vpermq \$0x15,$T2,$T2 # 00003412 -> 12343434
1595 vmovdqu `16*7-64`($ctx),%x#$D4
1596 vpermq \$0x15,$T3,$T3
1597 vpshufd \$0xc8,$T2,$T2 # 12343434 -> 14243444
1598 vmovdqu `16*8-64`($ctx),%x#$MASK
1599 vpermq \$0x15,$T4,$T4
1600 vpshufd \$0xc8,$T3,$T3
1601 vmovdqa $T2,0x00(%rsp)
1602 vpermq \$0x15,$D0,$D0
1603 vpshufd \$0xc8,$T4,$T4
1604 vmovdqa $T3,0x20(%rsp)
1605 vpermq \$0x15,$D1,$D1
1606 vpshufd \$0xc8,$D0,$D0
1607 vmovdqa $T4,0x40(%rsp)
1608 vpermq \$0x15,$D2,$D2
1609 vpshufd \$0xc8,$D1,$D1
1610 vmovdqa $D0,0x60(%rsp)
1611 vpermq \$0x15,$D3,$D3
1612 vpshufd \$0xc8,$D2,$D2
1613 vmovdqa $D1,0x80(%rsp)
1614 vpermq \$0x15,$D4,$D4
1615 vpshufd \$0xc8,$D3,$D3
1616 vmovdqa $D2,0xa0(%rsp)
1617 vpermq \$0x15,$MASK,$MASK
1618 vpshufd \$0xc8,$D4,$D4
1619 vmovdqa $D3,0xc0(%rsp)
1620 vpshufd \$0xc8,$MASK,$MASK
1621 vmovdqa $D4,0xe0(%rsp)
1622 vmovdqa $MASK,0x100(%rsp)
1623 vmovdqa 64(%rcx),$MASK # .Lmask26
1624
1625 ################################################################
1626 # load input
1627 vmovdqu 16*0($inp),%x#$T0
1628 vmovdqu 16*1($inp),%x#$T1
1629 vinserti128 \$1,16*2($inp),$T0,$T0
1630 vinserti128 \$1,16*3($inp),$T1,$T1
1631 lea 16*4($inp),$inp
1632
1633 vpsrldq \$6,$T0,$T2 # splat input
1634 vpsrldq \$6,$T1,$T3
1635 vpunpckhqdq $T1,$T0,$T4 # 4
1636 vpunpcklqdq $T3,$T2,$T2 # 2:3
1637 vpunpcklqdq $T1,$T0,$T0 # 0:1
1638
1639 vpsrlq \$30,$T2,$T3
1640 vpsrlq \$4,$T2,$T2
1641 vpsrlq \$26,$T0,$T1
1642 vpsrlq \$40,$T4,$T4 # 4
1643 vpand $MASK,$T2,$T2 # 2
1644 vpand $MASK,$T0,$T0 # 0
1645 vpand $MASK,$T1,$T1 # 1
1646 vpand $MASK,$T3,$T3 # 3
1647 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1648
1649 lea 0x90(%rsp),%rax # size optimization
1650 vpaddq $H2,$T2,$H2 # accumulate input
1651 sub \$64,$len
1652 jz .Ltail_avx2
1653 jmp .Loop_avx2
1654
1655 .align 32
1656 .Loop_avx2:
1657 ################################################################
1658 # ((inp[0]*r^4+r[4])*r^4+r[8])*r^4
1659 # ((inp[1]*r^4+r[5])*r^4+r[9])*r^3
1660 # ((inp[2]*r^4+r[6])*r^4+r[10])*r^2
1661 # ((inp[3]*r^4+r[7])*r^4+r[11])*r^1
1662 # \________/\________/
1663 ################################################################
1664 #vpaddq $H2,$T2,$H2 # accumulate input
1665 vpaddq $H0,$T0,$H0
1666 vmovdqa `32*0`(%rsp),$T0 # r0^4
1667 vpaddq $H1,$T1,$H1
1668 vmovdqa `32*1`(%rsp),$T1 # r1^4
1669 vpaddq $H3,$T3,$H3
1670 vmovdqa `32*3`(%rsp),$T2 # r2^4
1671 vpaddq $H4,$T4,$H4
1672 vmovdqa `32*6-0x90`(%rax),$T3 # s3^4
1673 vmovdqa `32*8-0x90`(%rax),$S4 # s4^4
1674
1675 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1676 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1677 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1678 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1679 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1680 #
1681 # however, as h2 is "chronologically" first one available pull
1682 # corresponding operations up, so it's
1683 #
1684 # d4 = h2*r2 + h4*r0 + h3*r1 + h1*r3 + h0*r4
1685 # d3 = h2*r1 + h3*r0 + h1*r2 + h0*r3 + h4*5*r4
1686 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1687 # d1 = h2*5*r4 + h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3
1688 # d0 = h2*5*r3 + h0*r0 + h4*5*r1 + h3*5*r2 + h1*5*r4
1689
1690 vpmuludq $H2,$T0,$D2 # d2 = h2*r0
1691 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
1692 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
1693 vpmuludq $H2,$T3,$D0 # d0 = h2*s3
1694 vpmuludq $H2,$S4,$D1 # d1 = h2*s4
1695
1696 vpmuludq $H0,$T1,$T4 # h0*r1
1697 vpmuludq $H1,$T1,$H2 # h1*r1, borrow $H2 as temp
1698 vpaddq $T4,$D1,$D1 # d1 += h0*r1
1699 vpaddq $H2,$D2,$D2 # d2 += h1*r1
1700 vpmuludq $H3,$T1,$T4 # h3*r1
1701 vpmuludq `32*2`(%rsp),$H4,$H2 # h4*s1
1702 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1703 vpaddq $H2,$D0,$D0 # d0 += h4*s1
1704 vmovdqa `32*4-0x90`(%rax),$T1 # s2
1705
1706 vpmuludq $H0,$T0,$T4 # h0*r0
1707 vpmuludq $H1,$T0,$H2 # h1*r0
1708 vpaddq $T4,$D0,$D0 # d0 += h0*r0
1709 vpaddq $H2,$D1,$D1 # d1 += h1*r0
1710 vpmuludq $H3,$T0,$T4 # h3*r0
1711 vpmuludq $H4,$T0,$H2 # h4*r0
1712 vmovdqu 16*0($inp),%x#$T0 # load input
1713 vpaddq $T4,$D3,$D3 # d3 += h3*r0
1714 vpaddq $H2,$D4,$D4 # d4 += h4*r0
1715 vinserti128 \$1,16*2($inp),$T0,$T0
1716
1717 vpmuludq $H3,$T1,$T4 # h3*s2
1718 vpmuludq $H4,$T1,$H2 # h4*s2
1719 vmovdqu 16*1($inp),%x#$T1
1720 vpaddq $T4,$D0,$D0 # d0 += h3*s2
1721 vpaddq $H2,$D1,$D1 # d1 += h4*s2
1722 vmovdqa `32*5-0x90`(%rax),$H2 # r3
1723 vpmuludq $H1,$T2,$T4 # h1*r2
1724 vpmuludq $H0,$T2,$T2 # h0*r2
1725 vpaddq $T4,$D3,$D3 # d3 += h1*r2
1726 vpaddq $T2,$D2,$D2 # d2 += h0*r2
1727 vinserti128 \$1,16*3($inp),$T1,$T1
1728 lea 16*4($inp),$inp
1729
1730 vpmuludq $H1,$H2,$T4 # h1*r3
1731 vpmuludq $H0,$H2,$H2 # h0*r3
1732 vpsrldq \$6,$T0,$T2 # splat input
1733 vpaddq $T4,$D4,$D4 # d4 += h1*r3
1734 vpaddq $H2,$D3,$D3 # d3 += h0*r3
1735 vpmuludq $H3,$T3,$T4 # h3*s3
1736 vpmuludq $H4,$T3,$H2 # h4*s3
1737 vpsrldq \$6,$T1,$T3
1738 vpaddq $T4,$D1,$D1 # d1 += h3*s3
1739 vpaddq $H2,$D2,$D2 # d2 += h4*s3
1740 vpunpckhqdq $T1,$T0,$T4 # 4
1741
1742 vpmuludq $H3,$S4,$H3 # h3*s4
1743 vpmuludq $H4,$S4,$H4 # h4*s4
1744 vpunpcklqdq $T1,$T0,$T0 # 0:1
1745 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*r4
1746 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*r4
1747 vpunpcklqdq $T3,$T2,$T3 # 2:3
1748 vpmuludq `32*7-0x90`(%rax),$H0,$H4 # h0*r4
1749 vpmuludq $H1,$S4,$H0 # h1*s4
1750 vmovdqa 64(%rcx),$MASK # .Lmask26
1751 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
1752 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
1753
1754 ################################################################
1755 # lazy reduction (interleaved with tail of input splat)
1756
1757 vpsrlq \$26,$H3,$D3
1758 vpand $MASK,$H3,$H3
1759 vpaddq $D3,$H4,$H4 # h3 -> h4
1760
1761 vpsrlq \$26,$H0,$D0
1762 vpand $MASK,$H0,$H0
1763 vpaddq $D0,$D1,$H1 # h0 -> h1
1764
1765 vpsrlq \$26,$H4,$D4
1766 vpand $MASK,$H4,$H4
1767
1768 vpsrlq \$4,$T3,$T2
1769
1770 vpsrlq \$26,$H1,$D1
1771 vpand $MASK,$H1,$H1
1772 vpaddq $D1,$H2,$H2 # h1 -> h2
1773
1774 vpaddq $D4,$H0,$H0
1775 vpsllq \$2,$D4,$D4
1776 vpaddq $D4,$H0,$H0 # h4 -> h0
1777
1778 vpand $MASK,$T2,$T2 # 2
1779 vpsrlq \$26,$T0,$T1
1780
1781 vpsrlq \$26,$H2,$D2
1782 vpand $MASK,$H2,$H2
1783 vpaddq $D2,$H3,$H3 # h2 -> h3
1784
1785 vpaddq $T2,$H2,$H2 # modulo-scheduled
1786 vpsrlq \$30,$T3,$T3
1787
1788 vpsrlq \$26,$H0,$D0
1789 vpand $MASK,$H0,$H0
1790 vpaddq $D0,$H1,$H1 # h0 -> h1
1791
1792 vpsrlq \$40,$T4,$T4 # 4
1793
1794 vpsrlq \$26,$H3,$D3
1795 vpand $MASK,$H3,$H3
1796 vpaddq $D3,$H4,$H4 # h3 -> h4
1797
1798 vpand $MASK,$T0,$T0 # 0
1799 vpand $MASK,$T1,$T1 # 1
1800 vpand $MASK,$T3,$T3 # 3
1801 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1802
1803 sub \$64,$len
1804 jnz .Loop_avx2
1805
1806 .byte 0x66,0x90
1807 .Ltail_avx2:
1808 ################################################################
1809 # while above multiplications were by r^4 in all lanes, in last
1810 # iteration we multiply least significant lane by r^4 and most
1811 # significant one by r, so copy of above except that references
1812 # to the precomputed table are displaced by 4...
1813
1814 #vpaddq $H2,$T2,$H2 # accumulate input
1815 vpaddq $H0,$T0,$H0
1816 vmovdqu `32*0+4`(%rsp),$T0 # r0^4
1817 vpaddq $H1,$T1,$H1
1818 vmovdqu `32*1+4`(%rsp),$T1 # r1^4
1819 vpaddq $H3,$T3,$H3
1820 vmovdqu `32*3+4`(%rsp),$T2 # r2^4
1821 vpaddq $H4,$T4,$H4
1822 vmovdqu `32*6+4-0x90`(%rax),$T3 # s3^4
1823 vmovdqu `32*8+4-0x90`(%rax),$S4 # s4^4
1824
1825 vpmuludq $H2,$T0,$D2 # d2 = h2*r0
1826 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
1827 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
1828 vpmuludq $H2,$T3,$D0 # d0 = h2*s3
1829 vpmuludq $H2,$S4,$D1 # d1 = h2*s4
1830
1831 vpmuludq $H0,$T1,$T4 # h0*r1
1832 vpmuludq $H1,$T1,$H2 # h1*r1
1833 vpaddq $T4,$D1,$D1 # d1 += h0*r1
1834 vpaddq $H2,$D2,$D2 # d2 += h1*r1
1835 vpmuludq $H3,$T1,$T4 # h3*r1
1836 vpmuludq `32*2+4`(%rsp),$H4,$H2 # h4*s1
1837 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1838 vpaddq $H2,$D0,$D0 # d0 += h4*s1
1839
1840 vpmuludq $H0,$T0,$T4 # h0*r0
1841 vpmuludq $H1,$T0,$H2 # h1*r0
1842 vpaddq $T4,$D0,$D0 # d0 += h0*r0
1843 vmovdqu `32*4+4-0x90`(%rax),$T1 # s2
1844 vpaddq $H2,$D1,$D1 # d1 += h1*r0
1845 vpmuludq $H3,$T0,$T4 # h3*r0
1846 vpmuludq $H4,$T0,$H2 # h4*r0
1847 vpaddq $T4,$D3,$D3 # d3 += h3*r0
1848 vpaddq $H2,$D4,$D4 # d4 += h4*r0
1849
1850 vpmuludq $H3,$T1,$T4 # h3*s2
1851 vpmuludq $H4,$T1,$H2 # h4*s2
1852 vpaddq $T4,$D0,$D0 # d0 += h3*s2
1853 vpaddq $H2,$D1,$D1 # d1 += h4*s2
1854 vmovdqu `32*5+4-0x90`(%rax),$H2 # r3
1855 vpmuludq $H1,$T2,$T4 # h1*r2
1856 vpmuludq $H0,$T2,$T2 # h0*r2
1857 vpaddq $T4,$D3,$D3 # d3 += h1*r2
1858 vpaddq $T2,$D2,$D2 # d2 += h0*r2
1859
1860 vpmuludq $H1,$H2,$T4 # h1*r3
1861 vpmuludq $H0,$H2,$H2 # h0*r3
1862 vpaddq $T4,$D4,$D4 # d4 += h1*r3
1863 vpaddq $H2,$D3,$D3 # d3 += h0*r3
1864 vpmuludq $H3,$T3,$T4 # h3*s3
1865 vpmuludq $H4,$T3,$H2 # h4*s3
1866 vpaddq $T4,$D1,$D1 # d1 += h3*s3
1867 vpaddq $H2,$D2,$D2 # d2 += h4*s3
1868
1869 vpmuludq $H3,$S4,$H3 # h3*s4
1870 vpmuludq $H4,$S4,$H4 # h4*s4
1871 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*r4
1872 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*r4
1873 vpmuludq `32*7+4-0x90`(%rax),$H0,$H4 # h0*r4
1874 vpmuludq $H1,$S4,$H0 # h1*s4
1875 vmovdqa 64(%rcx),$MASK # .Lmask26
1876 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
1877 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
1878
1879 ################################################################
1880 # horizontal addition
1881
1882 vpsrldq \$8,$D1,$T1
1883 vpsrldq \$8,$H2,$T2
1884 vpsrldq \$8,$H3,$T3
1885 vpsrldq \$8,$H4,$T4
1886 vpsrldq \$8,$H0,$T0
1887 vpaddq $T1,$D1,$D1
1888 vpaddq $T2,$H2,$H2
1889 vpaddq $T3,$H3,$H3
1890 vpaddq $T4,$H4,$H4
1891 vpaddq $T0,$H0,$H0
1892
1893 vpermq \$0x2,$H3,$T3
1894 vpermq \$0x2,$H4,$T4
1895 vpermq \$0x2,$H0,$T0
1896 vpermq \$0x2,$D1,$T1
1897 vpermq \$0x2,$H2,$T2
1898 vpaddq $T3,$H3,$H3
1899 vpaddq $T4,$H4,$H4
1900 vpaddq $T0,$H0,$H0
1901 vpaddq $T1,$D1,$D1
1902 vpaddq $T2,$H2,$H2
1903
1904 ################################################################
1905 # lazy reduction
1906
1907 vpsrlq \$26,$H3,$D3
1908 vpand $MASK,$H3,$H3
1909 vpaddq $D3,$H4,$H4 # h3 -> h4
1910
1911 vpsrlq \$26,$H0,$D0
1912 vpand $MASK,$H0,$H0
1913 vpaddq $D0,$D1,$H1 # h0 -> h1
1914
1915 vpsrlq \$26,$H4,$D4
1916 vpand $MASK,$H4,$H4
1917
1918 vpsrlq \$26,$H1,$D1
1919 vpand $MASK,$H1,$H1
1920 vpaddq $D1,$H2,$H2 # h1 -> h2
1921
1922 vpaddq $D4,$H0,$H0
1923 vpsllq \$2,$D4,$D4
1924 vpaddq $D4,$H0,$H0 # h4 -> h0
1925
1926 vpsrlq \$26,$H2,$D2
1927 vpand $MASK,$H2,$H2
1928 vpaddq $D2,$H3,$H3 # h2 -> h3
1929
1930 vpsrlq \$26,$H0,$D0
1931 vpand $MASK,$H0,$H0
1932 vpaddq $D0,$H1,$H1 # h0 -> h1
1933
1934 vpsrlq \$26,$H3,$D3
1935 vpand $MASK,$H3,$H3
1936 vpaddq $D3,$H4,$H4 # h3 -> h4
1937
1938 vmovd %x#$H0,`4*0-48-64`($ctx)# save partially reduced
1939 vmovd %x#$H1,`4*1-48-64`($ctx)
1940 vmovd %x#$H2,`4*2-48-64`($ctx)
1941 vmovd %x#$H3,`4*3-48-64`($ctx)
1942 vmovd %x#$H4,`4*4-48-64`($ctx)
1943 ___
1944 $code.=<<___ if ($win64);
1945 vmovdqa 0x50(%r11),%xmm6
1946 vmovdqa 0x60(%r11),%xmm7
1947 vmovdqa 0x70(%r11),%xmm8
1948 vmovdqa 0x80(%r11),%xmm9
1949 vmovdqa 0x90(%r11),%xmm10
1950 vmovdqa 0xa0(%r11),%xmm11
1951 vmovdqa 0xb0(%r11),%xmm12
1952 vmovdqa 0xc0(%r11),%xmm13
1953 vmovdqa 0xd0(%r11),%xmm14
1954 vmovdqa 0xe0(%r11),%xmm15
1955 lea 0xf8(%r11),%rsp
1956 .Ldo_avx2_epilogue:
1957 ___
1958 $code.=<<___ if (!$win64);
1959 lea 8(%r11),%rsp
1960 ___
1961 $code.=<<___;
1962 vzeroupper
1963 ret
1964 .size poly1305_blocks_avx2,.-poly1305_blocks_avx2
1965 ___
1966 }
1967 $code.=<<___;
1968 .align 64
1969 .Lconst:
1970 .Lmask24:
1971 .long 0x0ffffff,0,0x0ffffff,0,0x0ffffff,0,0x0ffffff,0
1972 .L129:
1973 .long 1<<24,0,1<<24,0,1<<24,0,1<<24,0
1974 .Lmask26:
1975 .long 0x3ffffff,0,0x3ffffff,0,0x3ffffff,0,0x3ffffff,0
1976 .Lfive:
1977 .long 5,0,5,0,5,0,5,0
1978 ___
1979 }
1980
1981 $code.=<<___;
1982 .asciz "Poly1305 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1983 .align 16
1984 ___
1985
1986 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1987 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1988 if ($win64) {
1989 $rec="%rcx";
1990 $frame="%rdx";
1991 $context="%r8";
1992 $disp="%r9";
1993
1994 $code.=<<___;
1995 .extern __imp_RtlVirtualUnwind
1996 .type se_handler,\@abi-omnipotent
1997 .align 16
1998 se_handler:
1999 push %rsi
2000 push %rdi
2001 push %rbx
2002 push %rbp
2003 push %r12
2004 push %r13
2005 push %r14
2006 push %r15
2007 pushfq
2008 sub \$64,%rsp
2009
2010 mov 120($context),%rax # pull context->Rax
2011 mov 248($context),%rbx # pull context->Rip
2012
2013 mov 8($disp),%rsi # disp->ImageBase
2014 mov 56($disp),%r11 # disp->HandlerData
2015
2016 mov 0(%r11),%r10d # HandlerData[0]
2017 lea (%rsi,%r10),%r10 # prologue label
2018 cmp %r10,%rbx # context->Rip<.Lprologue
2019 jb .Lcommon_seh_tail
2020
2021 mov 152($context),%rax # pull context->Rsp
2022
2023 mov 4(%r11),%r10d # HandlerData[1]
2024 lea (%rsi,%r10),%r10 # epilogue label
2025 cmp %r10,%rbx # context->Rip>=.Lepilogue
2026 jae .Lcommon_seh_tail
2027
2028 lea 48(%rax),%rax
2029
2030 mov -8(%rax),%rbx
2031 mov -16(%rax),%rbp
2032 mov -24(%rax),%r12
2033 mov -32(%rax),%r13
2034 mov -40(%rax),%r14
2035 mov -48(%rax),%r15
2036 mov %rbx,144($context) # restore context->Rbx
2037 mov %rbp,160($context) # restore context->Rbp
2038 mov %r12,216($context) # restore context->R12
2039 mov %r13,224($context) # restore context->R13
2040 mov %r14,232($context) # restore context->R14
2041 mov %r15,240($context) # restore context->R14
2042
2043 jmp .Lcommon_seh_tail
2044 .size se_handler,.-se_handler
2045
2046 .type avx_handler,\@abi-omnipotent
2047 .align 16
2048 avx_handler:
2049 push %rsi
2050 push %rdi
2051 push %rbx
2052 push %rbp
2053 push %r12
2054 push %r13
2055 push %r14
2056 push %r15
2057 pushfq
2058 sub \$64,%rsp
2059
2060 mov 120($context),%rax # pull context->Rax
2061 mov 248($context),%rbx # pull context->Rip
2062
2063 mov 8($disp),%rsi # disp->ImageBase
2064 mov 56($disp),%r11 # disp->HandlerData
2065
2066 mov 0(%r11),%r10d # HandlerData[0]
2067 lea (%rsi,%r10),%r10 # prologue label
2068 cmp %r10,%rbx # context->Rip<prologue label
2069 jb .Lcommon_seh_tail
2070
2071 mov 152($context),%rax # pull context->Rsp
2072
2073 mov 4(%r11),%r10d # HandlerData[1]
2074 lea (%rsi,%r10),%r10 # epilogue label
2075 cmp %r10,%rbx # context->Rip>=epilogue label
2076 jae .Lcommon_seh_tail
2077
2078 mov 208($context),%rax # pull context->R11
2079
2080 lea 0x50(%rax),%rsi
2081 lea 0xf8(%rax),%rax
2082 lea 512($context),%rdi # &context.Xmm6
2083 mov \$20,%ecx
2084 .long 0xa548f3fc # cld; rep movsq
2085
2086 .Lcommon_seh_tail:
2087 mov 8(%rax),%rdi
2088 mov 16(%rax),%rsi
2089 mov %rax,152($context) # restore context->Rsp
2090 mov %rsi,168($context) # restore context->Rsi
2091 mov %rdi,176($context) # restore context->Rdi
2092
2093 mov 40($disp),%rdi # disp->ContextRecord
2094 mov $context,%rsi # context
2095 mov \$154,%ecx # sizeof(CONTEXT)
2096 .long 0xa548f3fc # cld; rep movsq
2097
2098 mov $disp,%rsi
2099 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
2100 mov 8(%rsi),%rdx # arg2, disp->ImageBase
2101 mov 0(%rsi),%r8 # arg3, disp->ControlPc
2102 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
2103 mov 40(%rsi),%r10 # disp->ContextRecord
2104 lea 56(%rsi),%r11 # &disp->HandlerData
2105 lea 24(%rsi),%r12 # &disp->EstablisherFrame
2106 mov %r10,32(%rsp) # arg5
2107 mov %r11,40(%rsp) # arg6
2108 mov %r12,48(%rsp) # arg7
2109 mov %rcx,56(%rsp) # arg8, (NULL)
2110 call *__imp_RtlVirtualUnwind(%rip)
2111
2112 mov \$1,%eax # ExceptionContinueSearch
2113 add \$64,%rsp
2114 popfq
2115 pop %r15
2116 pop %r14
2117 pop %r13
2118 pop %r12
2119 pop %rbp
2120 pop %rbx
2121 pop %rdi
2122 pop %rsi
2123 ret
2124 .size avx_handler,.-avx_handler
2125
2126 .section .pdata
2127 .align 4
2128 .rva .LSEH_begin_poly1305_init
2129 .rva .LSEH_end_poly1305_init
2130 .rva .LSEH_info_poly1305_init
2131
2132 .rva .LSEH_begin_poly1305_blocks
2133 .rva .LSEH_end_poly1305_blocks
2134 .rva .LSEH_info_poly1305_blocks
2135
2136 .rva .LSEH_begin_poly1305_emit
2137 .rva .LSEH_end_poly1305_emit
2138 .rva .LSEH_info_poly1305_emit
2139 ___
2140 $code.=<<___ if ($avx);
2141 .rva .LSEH_begin_poly1305_blocks_avx
2142 .rva .Lbase2_64_avx
2143 .rva .LSEH_info_poly1305_blocks_avx_1
2144
2145 .rva .Lbase2_64_avx
2146 .rva .Leven_avx
2147 .rva .LSEH_info_poly1305_blocks_avx_2
2148
2149 .rva .Leven_avx
2150 .rva .LSEH_end_poly1305_blocks_avx
2151 .rva .LSEH_info_poly1305_blocks_avx_3
2152
2153 .rva .LSEH_begin_poly1305_emit_avx
2154 .rva .LSEH_end_poly1305_emit_avx
2155 .rva .LSEH_info_poly1305_emit_avx
2156 ___
2157 $code.=<<___ if ($avx>1);
2158 .rva .LSEH_begin_poly1305_blocks_avx2
2159 .rva .Lbase2_64_avx2
2160 .rva .LSEH_info_poly1305_blocks_avx2_1
2161
2162 .rva .Lbase2_64_avx2
2163 .rva .Leven_avx2
2164 .rva .LSEH_info_poly1305_blocks_avx2_2
2165
2166 .rva .Leven_avx2
2167 .rva .LSEH_end_poly1305_blocks_avx2
2168 .rva .LSEH_info_poly1305_blocks_avx2_3
2169 ___
2170 $code.=<<___;
2171 .section .xdata
2172 .align 8
2173 .LSEH_info_poly1305_init:
2174 .byte 9,0,0,0
2175 .rva se_handler
2176 .rva .LSEH_begin_poly1305_init,.LSEH_begin_poly1305_init
2177
2178 .LSEH_info_poly1305_blocks:
2179 .byte 9,0,0,0
2180 .rva se_handler
2181 .rva .Lblocks_body,.Lblocks_epilogue
2182
2183 .LSEH_info_poly1305_emit:
2184 .byte 9,0,0,0
2185 .rva se_handler
2186 .rva .LSEH_begin_poly1305_emit,.LSEH_begin_poly1305_emit
2187 ___
2188 $code.=<<___ if ($avx);
2189 .LSEH_info_poly1305_blocks_avx_1:
2190 .byte 9,0,0,0
2191 .rva se_handler
2192 .rva .Lblocks_avx_body,.Lblocks_avx_epilogue # HandlerData[]
2193
2194 .LSEH_info_poly1305_blocks_avx_2:
2195 .byte 9,0,0,0
2196 .rva se_handler
2197 .rva .Lbase2_64_avx_body,.Lbase2_64_avx_epilogue # HandlerData[]
2198
2199 .LSEH_info_poly1305_blocks_avx_3:
2200 .byte 9,0,0,0
2201 .rva avx_handler
2202 .rva .Ldo_avx_body,.Ldo_avx_epilogue # HandlerData[]
2203
2204 .LSEH_info_poly1305_emit_avx:
2205 .byte 9,0,0,0
2206 .rva se_handler
2207 .rva .LSEH_begin_poly1305_emit_avx,.LSEH_begin_poly1305_emit_avx
2208 ___
2209 $code.=<<___ if ($avx>1);
2210 .LSEH_info_poly1305_blocks_avx2_1:
2211 .byte 9,0,0,0
2212 .rva se_handler
2213 .rva .Lblocks_avx2_body,.Lblocks_avx2_epilogue # HandlerData[]
2214
2215 .LSEH_info_poly1305_blocks_avx2_2:
2216 .byte 9,0,0,0
2217 .rva se_handler
2218 .rva .Lbase2_64_avx2_body,.Lbase2_64_avx2_epilogue # HandlerData[]
2219
2220 .LSEH_info_poly1305_blocks_avx2_3:
2221 .byte 9,0,0,0
2222 .rva avx_handler
2223 .rva .Ldo_avx2_body,.Ldo_avx2_epilogue # HandlerData[]
2224 ___
2225 }
2226
2227 foreach (split('\n',$code)) {
2228 s/\`([^\`]*)\`/eval($1)/ge;
2229 s/%r([a-z]+)#d/%e$1/g;
2230 s/%r([0-9]+)#d/%r$1d/g;
2231 s/%x#%y/%x/g;
2232
2233 print $_,"\n";
2234 }
2235 close STDOUT;