xref: /aosp_15_r20/external/boringssl/src/crypto/chacha/asm/chacha-x86_64.pl (revision 8fb009dc861624b67b6cdb62ea21f0f22d0c584b) 
1#! /usr/bin/env perl
2# Copyright 2016 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9#
10# ====================================================================
11# Written by Andy Polyakov <[email protected]> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16#
17# November 2014
18#
19# ChaCha20 for x86_64.
20#
21# December 2016
22#
23# Add AVX512F code path.
24#
25# Performance in cycles per byte out of large buffer.
26#
27#		IALU/gcc 4.8(i)	1xSSSE3/SSE2	4xSSSE3	    NxAVX(v)
28#
29# P4		9.48/+99%	-/22.7(ii)	-
30# Core2		7.83/+55%	7.90/8.08	4.35
31# Westmere	7.19/+50%	5.60/6.70	3.00
32# Sandy Bridge	8.31/+42%	5.45/6.76	2.72
33# Ivy Bridge	6.71/+46%	5.40/6.49	2.41
34# Haswell	5.92/+43%	5.20/6.45	2.42	    1.23
35# Skylake[-X]	5.87/+39%	4.70/-		2.31	    1.19[0.57]
36# Silvermont	12.0/+33%	7.75/7.40	7.03(iii)
37# Knights L	11.7/-		-		9.60(iii)   0.80
38# Goldmont	10.6/+17%	5.10/-		3.28
39# Sledgehammer	7.28/+52%	-/14.2(ii)	-
40# Bulldozer	9.66/+28%	9.85/11.1	3.06(iv)
41# Ryzen		5.96/+50%	5.19/-		2.40        2.09
42# VIA Nano	10.5/+46%	6.72/8.60	6.05
43#
44# (i)	compared to older gcc 3.x one can observe >2x improvement on
45#	most platforms;
46# (ii)	as it can be seen, SSE2 performance is too low on legacy
47#	processors; NxSSE2 results are naturally better, but not
48#	impressively better than IALU ones, which is why you won't
49#	find SSE2 code below;
50# (iii)	this is not optimal result for Atom because of MSROM
51#	limitations, SSE2 can do better, but gain is considered too
52#	low to justify the [maintenance] effort;
53# (iv)	Bulldozer actually executes 4xXOP code path that delivers 2.20;
54#
55# Modified from upstream OpenSSL to remove the XOP code.
56
57$flavour = shift;
58$output  = shift;
59if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
60
61$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
62
63$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
64( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
65( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
66die "can't locate x86_64-xlate.pl";
67
68$avx = 2;
69
70open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
71*STDOUT=*OUT;
72
73# input parameter block
74($out,$inp,$len,$key,$counter)=("%rdi","%rsi","%rdx","%rcx","%r8");
75
76$code.=<<___;
77.text
78
79.section .rodata
80.align	64
81.Lzero:
82.long	0,0,0,0
83.Lone:
84.long	1,0,0,0
85.Linc:
86.long	0,1,2,3
87.Lfour:
88.long	4,4,4,4
89.Lincy:
90.long	0,2,4,6,1,3,5,7
91.Leight:
92.long	8,8,8,8,8,8,8,8
93.Lrot16:
94.byte	0x2,0x3,0x0,0x1, 0x6,0x7,0x4,0x5, 0xa,0xb,0x8,0x9, 0xe,0xf,0xc,0xd
95.Lrot24:
96.byte	0x3,0x0,0x1,0x2, 0x7,0x4,0x5,0x6, 0xb,0x8,0x9,0xa, 0xf,0xc,0xd,0xe
97.Lsigma:
98.asciz	"expand 32-byte k"
99.align	64
100.Lzeroz:
101.long	0,0,0,0, 1,0,0,0, 2,0,0,0, 3,0,0,0
102.Lfourz:
103.long	4,0,0,0, 4,0,0,0, 4,0,0,0, 4,0,0,0
104.Lincz:
105.long	0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
106.Lsixteen:
107.long	16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16
108.asciz	"ChaCha20 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
109.text
110___
111
112sub AUTOLOAD()          # thunk [simplified] 32-bit style perlasm
113{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
114  my $arg = pop;
115    $arg = "\$$arg" if ($arg*1 eq $arg);
116    $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
117}
118
119@x=("%eax","%ebx","%ecx","%edx",map("%r${_}d",(8..11)),
120    "%nox","%nox","%nox","%nox",map("%r${_}d",(12..15)));
121@t=("%esi","%edi");
122
123sub ROUND {			# critical path is 24 cycles per round
124my ($a0,$b0,$c0,$d0)=@_;
125my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
126my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
127my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
128my ($xc,$xc_)=map("\"$_\"",@t);
129my @x=map("\"$_\"",@x);
130
131	# Consider order in which variables are addressed by their
132	# index:
133	#
134	#	a   b   c   d
135	#
136	#	0   4   8  12 < even round
137	#	1   5   9  13
138	#	2   6  10  14
139	#	3   7  11  15
140	#	0   5  10  15 < odd round
141	#	1   6  11  12
142	#	2   7   8  13
143	#	3   4   9  14
144	#
145	# 'a', 'b' and 'd's are permanently allocated in registers,
146	# @x[0..7,12..15], while 'c's are maintained in memory. If
147	# you observe 'c' column, you'll notice that pair of 'c's is
148	# invariant between rounds. This means that we have to reload
149	# them once per round, in the middle. This is why you'll see
150	# bunch of 'c' stores and loads in the middle, but none in
151	# the beginning or end.
152
153	# Normally instructions would be interleaved to favour in-order
154	# execution. Generally out-of-order cores manage it gracefully,
155	# but not this time for some reason. As in-order execution
156	# cores are dying breed, old Atom is the only one around,
157	# instructions are left uninterleaved. Besides, Atom is better
158	# off executing 1xSSSE3 code anyway...
159
160	(
161	"&add	(@x[$a0],@x[$b0])",	# Q1
162	"&xor	(@x[$d0],@x[$a0])",
163	"&rol	(@x[$d0],16)",
164	 "&add	(@x[$a1],@x[$b1])",	# Q2
165	 "&xor	(@x[$d1],@x[$a1])",
166	 "&rol	(@x[$d1],16)",
167
168	"&add	($xc,@x[$d0])",
169	"&xor	(@x[$b0],$xc)",
170	"&rol	(@x[$b0],12)",
171	 "&add	($xc_,@x[$d1])",
172	 "&xor	(@x[$b1],$xc_)",
173	 "&rol	(@x[$b1],12)",
174
175	"&add	(@x[$a0],@x[$b0])",
176	"&xor	(@x[$d0],@x[$a0])",
177	"&rol	(@x[$d0],8)",
178	 "&add	(@x[$a1],@x[$b1])",
179	 "&xor	(@x[$d1],@x[$a1])",
180	 "&rol	(@x[$d1],8)",
181
182	"&add	($xc,@x[$d0])",
183	"&xor	(@x[$b0],$xc)",
184	"&rol	(@x[$b0],7)",
185	 "&add	($xc_,@x[$d1])",
186	 "&xor	(@x[$b1],$xc_)",
187	 "&rol	(@x[$b1],7)",
188
189	"&mov	(\"4*$c0(%rsp)\",$xc)",	# reload pair of 'c's
190	 "&mov	(\"4*$c1(%rsp)\",$xc_)",
191	"&mov	($xc,\"4*$c2(%rsp)\")",
192	 "&mov	($xc_,\"4*$c3(%rsp)\")",
193
194	"&add	(@x[$a2],@x[$b2])",	# Q3
195	"&xor	(@x[$d2],@x[$a2])",
196	"&rol	(@x[$d2],16)",
197	 "&add	(@x[$a3],@x[$b3])",	# Q4
198	 "&xor	(@x[$d3],@x[$a3])",
199	 "&rol	(@x[$d3],16)",
200
201	"&add	($xc,@x[$d2])",
202	"&xor	(@x[$b2],$xc)",
203	"&rol	(@x[$b2],12)",
204	 "&add	($xc_,@x[$d3])",
205	 "&xor	(@x[$b3],$xc_)",
206	 "&rol	(@x[$b3],12)",
207
208	"&add	(@x[$a2],@x[$b2])",
209	"&xor	(@x[$d2],@x[$a2])",
210	"&rol	(@x[$d2],8)",
211	 "&add	(@x[$a3],@x[$b3])",
212	 "&xor	(@x[$d3],@x[$a3])",
213	 "&rol	(@x[$d3],8)",
214
215	"&add	($xc,@x[$d2])",
216	"&xor	(@x[$b2],$xc)",
217	"&rol	(@x[$b2],7)",
218	 "&add	($xc_,@x[$d3])",
219	 "&xor	(@x[$b3],$xc_)",
220	 "&rol	(@x[$b3],7)"
221	);
222}
223
224########################################################################
225# Generic code path that handles all lengths on pre-SSSE3 processors.
226$code.=<<___;
227.globl	ChaCha20_ctr32_nohw
228.type	ChaCha20_ctr32_nohw,\@function,5
229.align	64
230ChaCha20_ctr32_nohw:
231.cfi_startproc
232	_CET_ENDBR
233	push	%rbx
234.cfi_push	rbx
235	push	%rbp
236.cfi_push	rbp
237	push	%r12
238.cfi_push	r12
239	push	%r13
240.cfi_push	r13
241	push	%r14
242.cfi_push	r14
243	push	%r15
244.cfi_push	r15
245	sub	\$64+24,%rsp
246.cfi_adjust_cfa_offset	`64+24`
247.Lctr32_body:
248
249	#movdqa	.Lsigma(%rip),%xmm0
250	movdqu	($key),%xmm1
251	movdqu	16($key),%xmm2
252	movdqu	($counter),%xmm3
253	movdqa	.Lone(%rip),%xmm4
254
255	#movdqa	%xmm0,4*0(%rsp)		# key[0]
256	movdqa	%xmm1,4*4(%rsp)		# key[1]
257	movdqa	%xmm2,4*8(%rsp)		# key[2]
258	movdqa	%xmm3,4*12(%rsp)	# key[3]
259	mov	$len,%rbp		# reassign $len
260	jmp	.Loop_outer
261
262.align	32
263.Loop_outer:
264	mov	\$0x61707865,@x[0]      # 'expa'
265	mov	\$0x3320646e,@x[1]      # 'nd 3'
266	mov	\$0x79622d32,@x[2]      # '2-by'
267	mov	\$0x6b206574,@x[3]      # 'te k'
268	mov	4*4(%rsp),@x[4]
269	mov	4*5(%rsp),@x[5]
270	mov	4*6(%rsp),@x[6]
271	mov	4*7(%rsp),@x[7]
272	movd	%xmm3,@x[12]
273	mov	4*13(%rsp),@x[13]
274	mov	4*14(%rsp),@x[14]
275	mov	4*15(%rsp),@x[15]
276
277	mov	%rbp,64+0(%rsp)		# save len
278	mov	\$10,%ebp
279	mov	$inp,64+8(%rsp)		# save inp
280	movq	%xmm2,%rsi		# "@x[8]"
281	mov	$out,64+16(%rsp)	# save out
282	mov	%rsi,%rdi
283	shr	\$32,%rdi		# "@x[9]"
284	jmp	.Loop
285
286.align	32
287.Loop:
288___
289	foreach (&ROUND (0, 4, 8,12)) { eval; }
290	foreach (&ROUND	(0, 5,10,15)) { eval; }
291	&dec	("%ebp");
292	&jnz	(".Loop");
293
294$code.=<<___;
295	mov	@t[1],4*9(%rsp)		# modulo-scheduled
296	mov	@t[0],4*8(%rsp)
297	mov	64(%rsp),%rbp		# load len
298	movdqa	%xmm2,%xmm1
299	mov	64+8(%rsp),$inp		# load inp
300	paddd	%xmm4,%xmm3		# increment counter
301	mov	64+16(%rsp),$out	# load out
302
303	add	\$0x61707865,@x[0]      # 'expa'
304	add	\$0x3320646e,@x[1]      # 'nd 3'
305	add	\$0x79622d32,@x[2]      # '2-by'
306	add	\$0x6b206574,@x[3]      # 'te k'
307	add	4*4(%rsp),@x[4]
308	add	4*5(%rsp),@x[5]
309	add	4*6(%rsp),@x[6]
310	add	4*7(%rsp),@x[7]
311	add	4*12(%rsp),@x[12]
312	add	4*13(%rsp),@x[13]
313	add	4*14(%rsp),@x[14]
314	add	4*15(%rsp),@x[15]
315	paddd	4*8(%rsp),%xmm1
316
317	cmp	\$64,%rbp
318	jb	.Ltail
319
320	xor	4*0($inp),@x[0]		# xor with input
321	xor	4*1($inp),@x[1]
322	xor	4*2($inp),@x[2]
323	xor	4*3($inp),@x[3]
324	xor	4*4($inp),@x[4]
325	xor	4*5($inp),@x[5]
326	xor	4*6($inp),@x[6]
327	xor	4*7($inp),@x[7]
328	movdqu	4*8($inp),%xmm0
329	xor	4*12($inp),@x[12]
330	xor	4*13($inp),@x[13]
331	xor	4*14($inp),@x[14]
332	xor	4*15($inp),@x[15]
333	lea	4*16($inp),$inp		# inp+=64
334	pxor	%xmm1,%xmm0
335
336	movdqa	%xmm2,4*8(%rsp)
337	movd	%xmm3,4*12(%rsp)
338
339	mov	@x[0],4*0($out)		# write output
340	mov	@x[1],4*1($out)
341	mov	@x[2],4*2($out)
342	mov	@x[3],4*3($out)
343	mov	@x[4],4*4($out)
344	mov	@x[5],4*5($out)
345	mov	@x[6],4*6($out)
346	mov	@x[7],4*7($out)
347	movdqu	%xmm0,4*8($out)
348	mov	@x[12],4*12($out)
349	mov	@x[13],4*13($out)
350	mov	@x[14],4*14($out)
351	mov	@x[15],4*15($out)
352	lea	4*16($out),$out		# out+=64
353
354	sub	\$64,%rbp
355	jnz	.Loop_outer
356
357	jmp	.Ldone
358
359.align	16
360.Ltail:
361	mov	@x[0],4*0(%rsp)
362	mov	@x[1],4*1(%rsp)
363	xor	%rbx,%rbx
364	mov	@x[2],4*2(%rsp)
365	mov	@x[3],4*3(%rsp)
366	mov	@x[4],4*4(%rsp)
367	mov	@x[5],4*5(%rsp)
368	mov	@x[6],4*6(%rsp)
369	mov	@x[7],4*7(%rsp)
370	movdqa	%xmm1,4*8(%rsp)
371	mov	@x[12],4*12(%rsp)
372	mov	@x[13],4*13(%rsp)
373	mov	@x[14],4*14(%rsp)
374	mov	@x[15],4*15(%rsp)
375
376.Loop_tail:
377	movzb	($inp,%rbx),%eax
378	movzb	(%rsp,%rbx),%edx
379	lea	1(%rbx),%rbx
380	xor	%edx,%eax
381	mov	%al,-1($out,%rbx)
382	dec	%rbp
383	jnz	.Loop_tail
384
385.Ldone:
386	lea	64+24+48(%rsp),%rsi
387	mov	-48(%rsi),%r15
388.cfi_restore	r15
389	mov	-40(%rsi),%r14
390.cfi_restore	r14
391	mov	-32(%rsi),%r13
392.cfi_restore	r13
393	mov	-24(%rsi),%r12
394.cfi_restore	r12
395	mov	-16(%rsi),%rbp
396.cfi_restore	rbp
397	mov	-8(%rsi),%rbx
398.cfi_restore	rbx
399	lea	(%rsi),%rsp
400.cfi_adjust_cfa_offset	`-64-24-48`
401.Lno_data:
402	ret
403.cfi_endproc
404.size	ChaCha20_ctr32_nohw,.-ChaCha20_ctr32_nohw
405___
406
407########################################################################
408# SSSE3 code path that handles shorter lengths
409{
410my ($a,$b,$c,$d,$t,$t1,$rot16,$rot24)=map("%xmm$_",(0..7));
411
412sub SSSE3ROUND {	# critical path is 20 "SIMD ticks" per round
413	&paddd	($a,$b);
414	&pxor	($d,$a);
415	&pshufb	($d,$rot16);
416
417	&paddd	($c,$d);
418	&pxor	($b,$c);
419	&movdqa	($t,$b);
420	&psrld	($b,20);
421	&pslld	($t,12);
422	&por	($b,$t);
423
424	&paddd	($a,$b);
425	&pxor	($d,$a);
426	&pshufb	($d,$rot24);
427
428	&paddd	($c,$d);
429	&pxor	($b,$c);
430	&movdqa	($t,$b);
431	&psrld	($b,25);
432	&pslld	($t,7);
433	&por	($b,$t);
434}
435
436my $xframe = $win64 ? 32+8 : 8;
437
438$code.=<<___;
439.globl	ChaCha20_ctr32_ssse3
440.type	ChaCha20_ctr32_ssse3,\@function,5
441.align	32
442ChaCha20_ctr32_ssse3:
443.cfi_startproc
444	_CET_ENDBR
445	mov	%rsp,%r9		# frame pointer
446.cfi_def_cfa_register	r9
447___
448$code.=<<___;
449	sub	\$64+$xframe,%rsp
450___
451$code.=<<___	if ($win64);
452	movaps	%xmm6,-0x28(%r9)
453	movaps	%xmm7,-0x18(%r9)
454.Lssse3_body:
455___
456$code.=<<___;
457	movdqa	.Lsigma(%rip),$a
458	movdqu	($key),$b
459	movdqu	16($key),$c
460	movdqu	($counter),$d
461	movdqa	.Lrot16(%rip),$rot16
462	movdqa	.Lrot24(%rip),$rot24
463
464	movdqa	$a,0x00(%rsp)
465	movdqa	$b,0x10(%rsp)
466	movdqa	$c,0x20(%rsp)
467	movdqa	$d,0x30(%rsp)
468	mov	\$10,$counter		# reuse $counter
469	jmp	.Loop_ssse3
470
471.align	32
472.Loop_outer_ssse3:
473	movdqa	.Lone(%rip),$d
474	movdqa	0x00(%rsp),$a
475	movdqa	0x10(%rsp),$b
476	movdqa	0x20(%rsp),$c
477	paddd	0x30(%rsp),$d
478	mov	\$10,$counter
479	movdqa	$d,0x30(%rsp)
480	jmp	.Loop_ssse3
481
482.align	32
483.Loop_ssse3:
484___
485	&SSSE3ROUND();
486	&pshufd	($c,$c,0b01001110);
487	&pshufd	($b,$b,0b00111001);
488	&pshufd	($d,$d,0b10010011);
489	&nop	();
490
491	&SSSE3ROUND();
492	&pshufd	($c,$c,0b01001110);
493	&pshufd	($b,$b,0b10010011);
494	&pshufd	($d,$d,0b00111001);
495
496	&dec	($counter);
497	&jnz	(".Loop_ssse3");
498
499$code.=<<___;
500	paddd	0x00(%rsp),$a
501	paddd	0x10(%rsp),$b
502	paddd	0x20(%rsp),$c
503	paddd	0x30(%rsp),$d
504
505	cmp	\$64,$len
506	jb	.Ltail_ssse3
507
508	movdqu	0x00($inp),$t
509	movdqu	0x10($inp),$t1
510	pxor	$t,$a			# xor with input
511	movdqu	0x20($inp),$t
512	pxor	$t1,$b
513	movdqu	0x30($inp),$t1
514	lea	0x40($inp),$inp		# inp+=64
515	pxor	$t,$c
516	pxor	$t1,$d
517
518	movdqu	$a,0x00($out)		# write output
519	movdqu	$b,0x10($out)
520	movdqu	$c,0x20($out)
521	movdqu	$d,0x30($out)
522	lea	0x40($out),$out		# out+=64
523
524	sub	\$64,$len
525	jnz	.Loop_outer_ssse3
526
527	jmp	.Ldone_ssse3
528
529.align	16
530.Ltail_ssse3:
531	movdqa	$a,0x00(%rsp)
532	movdqa	$b,0x10(%rsp)
533	movdqa	$c,0x20(%rsp)
534	movdqa	$d,0x30(%rsp)
535	xor	$counter,$counter
536
537.Loop_tail_ssse3:
538	movzb	($inp,$counter),%eax
539	movzb	(%rsp,$counter),%ecx
540	lea	1($counter),$counter
541	xor	%ecx,%eax
542	mov	%al,-1($out,$counter)
543	dec	$len
544	jnz	.Loop_tail_ssse3
545
546.Ldone_ssse3:
547___
548$code.=<<___	if ($win64);
549	movaps	-0x28(%r9),%xmm6
550	movaps	-0x18(%r9),%xmm7
551___
552$code.=<<___;
553	lea	(%r9),%rsp
554.cfi_def_cfa_register	rsp
555.Lssse3_epilogue:
556	ret
557.cfi_endproc
558.size	ChaCha20_ctr32_ssse3,.-ChaCha20_ctr32_ssse3
559___
560}
561
562########################################################################
563# SSSE3 code path that handles longer messages.
564{
565# assign variables to favor Atom front-end
566my ($xd0,$xd1,$xd2,$xd3, $xt0,$xt1,$xt2,$xt3,
567    $xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3)=map("%xmm$_",(0..15));
568my  @xx=($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
569	"%nox","%nox","%nox","%nox", $xd0,$xd1,$xd2,$xd3);
570
571sub SSSE3_lane_ROUND {
572my ($a0,$b0,$c0,$d0)=@_;
573my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
574my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
575my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
576my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
577my @x=map("\"$_\"",@xx);
578
579	# Consider order in which variables are addressed by their
580	# index:
581	#
582	#	a   b   c   d
583	#
584	#	0   4   8  12 < even round
585	#	1   5   9  13
586	#	2   6  10  14
587	#	3   7  11  15
588	#	0   5  10  15 < odd round
589	#	1   6  11  12
590	#	2   7   8  13
591	#	3   4   9  14
592	#
593	# 'a', 'b' and 'd's are permanently allocated in registers,
594	# @x[0..7,12..15], while 'c's are maintained in memory. If
595	# you observe 'c' column, you'll notice that pair of 'c's is
596	# invariant between rounds. This means that we have to reload
597	# them once per round, in the middle. This is why you'll see
598	# bunch of 'c' stores and loads in the middle, but none in
599	# the beginning or end.
600
601	(
602	"&paddd		(@x[$a0],@x[$b0])",	# Q1
603	 "&paddd	(@x[$a1],@x[$b1])",	# Q2
604	"&pxor		(@x[$d0],@x[$a0])",
605	 "&pxor		(@x[$d1],@x[$a1])",
606	"&pshufb	(@x[$d0],$t1)",
607	 "&pshufb	(@x[$d1],$t1)",
608
609	"&paddd		($xc,@x[$d0])",
610	 "&paddd	($xc_,@x[$d1])",
611	"&pxor		(@x[$b0],$xc)",
612	 "&pxor		(@x[$b1],$xc_)",
613	"&movdqa	($t0,@x[$b0])",
614	"&pslld		(@x[$b0],12)",
615	"&psrld		($t0,20)",
616	 "&movdqa	($t1,@x[$b1])",
617	 "&pslld	(@x[$b1],12)",
618	"&por		(@x[$b0],$t0)",
619	 "&psrld	($t1,20)",
620	"&movdqa	($t0,'(%r11)')",	# .Lrot24(%rip)
621	 "&por		(@x[$b1],$t1)",
622
623	"&paddd		(@x[$a0],@x[$b0])",
624	 "&paddd	(@x[$a1],@x[$b1])",
625	"&pxor		(@x[$d0],@x[$a0])",
626	 "&pxor		(@x[$d1],@x[$a1])",
627	"&pshufb	(@x[$d0],$t0)",
628	 "&pshufb	(@x[$d1],$t0)",
629
630	"&paddd		($xc,@x[$d0])",
631	 "&paddd	($xc_,@x[$d1])",
632	"&pxor		(@x[$b0],$xc)",
633	 "&pxor		(@x[$b1],$xc_)",
634	"&movdqa	($t1,@x[$b0])",
635	"&pslld		(@x[$b0],7)",
636	"&psrld		($t1,25)",
637	 "&movdqa	($t0,@x[$b1])",
638	 "&pslld	(@x[$b1],7)",
639	"&por		(@x[$b0],$t1)",
640	 "&psrld	($t0,25)",
641	"&movdqa	($t1,'(%r10)')",	# .Lrot16(%rip)
642	 "&por		(@x[$b1],$t0)",
643
644	"&movdqa	(\"`16*($c0-8)`(%rsp)\",$xc)",	# reload pair of 'c's
645	 "&movdqa	(\"`16*($c1-8)`(%rsp)\",$xc_)",
646	"&movdqa	($xc,\"`16*($c2-8)`(%rsp)\")",
647	 "&movdqa	($xc_,\"`16*($c3-8)`(%rsp)\")",
648
649	"&paddd		(@x[$a2],@x[$b2])",	# Q3
650	 "&paddd	(@x[$a3],@x[$b3])",	# Q4
651	"&pxor		(@x[$d2],@x[$a2])",
652	 "&pxor		(@x[$d3],@x[$a3])",
653	"&pshufb	(@x[$d2],$t1)",
654	 "&pshufb	(@x[$d3],$t1)",
655
656	"&paddd		($xc,@x[$d2])",
657	 "&paddd	($xc_,@x[$d3])",
658	"&pxor		(@x[$b2],$xc)",
659	 "&pxor		(@x[$b3],$xc_)",
660	"&movdqa	($t0,@x[$b2])",
661	"&pslld		(@x[$b2],12)",
662	"&psrld		($t0,20)",
663	 "&movdqa	($t1,@x[$b3])",
664	 "&pslld	(@x[$b3],12)",
665	"&por		(@x[$b2],$t0)",
666	 "&psrld	($t1,20)",
667	"&movdqa	($t0,'(%r11)')",	# .Lrot24(%rip)
668	 "&por		(@x[$b3],$t1)",
669
670	"&paddd		(@x[$a2],@x[$b2])",
671	 "&paddd	(@x[$a3],@x[$b3])",
672	"&pxor		(@x[$d2],@x[$a2])",
673	 "&pxor		(@x[$d3],@x[$a3])",
674	"&pshufb	(@x[$d2],$t0)",
675	 "&pshufb	(@x[$d3],$t0)",
676
677	"&paddd		($xc,@x[$d2])",
678	 "&paddd	($xc_,@x[$d3])",
679	"&pxor		(@x[$b2],$xc)",
680	 "&pxor		(@x[$b3],$xc_)",
681	"&movdqa	($t1,@x[$b2])",
682	"&pslld		(@x[$b2],7)",
683	"&psrld		($t1,25)",
684	 "&movdqa	($t0,@x[$b3])",
685	 "&pslld	(@x[$b3],7)",
686	"&por		(@x[$b2],$t1)",
687	 "&psrld	($t0,25)",
688	"&movdqa	($t1,'(%r10)')",	# .Lrot16(%rip)
689	 "&por		(@x[$b3],$t0)"
690	);
691}
692
693my $xframe = $win64 ? 0xa8 : 8;
694
695$code.=<<___;
696.globl	ChaCha20_ctr32_ssse3_4x
697.type	ChaCha20_ctr32_ssse3_4x,\@function,5
698.align	32
699ChaCha20_ctr32_ssse3_4x:
700.cfi_startproc
701	_CET_ENDBR
702	mov		%rsp,%r9		# frame pointer
703.cfi_def_cfa_register	r9
704___
705$code.=<<___;
706	sub		\$0x140+$xframe,%rsp
707___
708	################ stack layout
709	# +0x00		SIMD equivalent of @x[8-12]
710	# ...
711	# +0x40		constant copy of key[0-2] smashed by lanes
712	# ...
713	# +0x100	SIMD counters (with nonce smashed by lanes)
714	# ...
715	# +0x140
716$code.=<<___	if ($win64);
717	movaps		%xmm6,-0xa8(%r9)
718	movaps		%xmm7,-0x98(%r9)
719	movaps		%xmm8,-0x88(%r9)
720	movaps		%xmm9,-0x78(%r9)
721	movaps		%xmm10,-0x68(%r9)
722	movaps		%xmm11,-0x58(%r9)
723	movaps		%xmm12,-0x48(%r9)
724	movaps		%xmm13,-0x38(%r9)
725	movaps		%xmm14,-0x28(%r9)
726	movaps		%xmm15,-0x18(%r9)
727.L4x_body:
728___
729$code.=<<___;
730	movdqa		.Lsigma(%rip),$xa3	# key[0]
731	movdqu		($key),$xb3		# key[1]
732	movdqu		16($key),$xt3		# key[2]
733	movdqu		($counter),$xd3		# key[3]
734	lea		0x100(%rsp),%rcx	# size optimization
735	lea		.Lrot16(%rip),%r10
736	lea		.Lrot24(%rip),%r11
737
738	pshufd		\$0x00,$xa3,$xa0	# smash key by lanes...
739	pshufd		\$0x55,$xa3,$xa1
740	movdqa		$xa0,0x40(%rsp)		# ... and offload
741	pshufd		\$0xaa,$xa3,$xa2
742	movdqa		$xa1,0x50(%rsp)
743	pshufd		\$0xff,$xa3,$xa3
744	movdqa		$xa2,0x60(%rsp)
745	movdqa		$xa3,0x70(%rsp)
746
747	pshufd		\$0x00,$xb3,$xb0
748	pshufd		\$0x55,$xb3,$xb1
749	movdqa		$xb0,0x80-0x100(%rcx)
750	pshufd		\$0xaa,$xb3,$xb2
751	movdqa		$xb1,0x90-0x100(%rcx)
752	pshufd		\$0xff,$xb3,$xb3
753	movdqa		$xb2,0xa0-0x100(%rcx)
754	movdqa		$xb3,0xb0-0x100(%rcx)
755
756	pshufd		\$0x00,$xt3,$xt0	# "$xc0"
757	pshufd		\$0x55,$xt3,$xt1	# "$xc1"
758	movdqa		$xt0,0xc0-0x100(%rcx)
759	pshufd		\$0xaa,$xt3,$xt2	# "$xc2"
760	movdqa		$xt1,0xd0-0x100(%rcx)
761	pshufd		\$0xff,$xt3,$xt3	# "$xc3"
762	movdqa		$xt2,0xe0-0x100(%rcx)
763	movdqa		$xt3,0xf0-0x100(%rcx)
764
765	pshufd		\$0x00,$xd3,$xd0
766	pshufd		\$0x55,$xd3,$xd1
767	paddd		.Linc(%rip),$xd0	# don't save counters yet
768	pshufd		\$0xaa,$xd3,$xd2
769	movdqa		$xd1,0x110-0x100(%rcx)
770	pshufd		\$0xff,$xd3,$xd3
771	movdqa		$xd2,0x120-0x100(%rcx)
772	movdqa		$xd3,0x130-0x100(%rcx)
773
774	jmp		.Loop_enter4x
775
776.align	32
777.Loop_outer4x:
778	movdqa		0x40(%rsp),$xa0		# re-load smashed key
779	movdqa		0x50(%rsp),$xa1
780	movdqa		0x60(%rsp),$xa2
781	movdqa		0x70(%rsp),$xa3
782	movdqa		0x80-0x100(%rcx),$xb0
783	movdqa		0x90-0x100(%rcx),$xb1
784	movdqa		0xa0-0x100(%rcx),$xb2
785	movdqa		0xb0-0x100(%rcx),$xb3
786	movdqa		0xc0-0x100(%rcx),$xt0	# "$xc0"
787	movdqa		0xd0-0x100(%rcx),$xt1	# "$xc1"
788	movdqa		0xe0-0x100(%rcx),$xt2	# "$xc2"
789	movdqa		0xf0-0x100(%rcx),$xt3	# "$xc3"
790	movdqa		0x100-0x100(%rcx),$xd0
791	movdqa		0x110-0x100(%rcx),$xd1
792	movdqa		0x120-0x100(%rcx),$xd2
793	movdqa		0x130-0x100(%rcx),$xd3
794	paddd		.Lfour(%rip),$xd0	# next SIMD counters
795
796.Loop_enter4x:
797	movdqa		$xt2,0x20(%rsp)		# SIMD equivalent of "@x[10]"
798	movdqa		$xt3,0x30(%rsp)		# SIMD equivalent of "@x[11]"
799	movdqa		(%r10),$xt3		# .Lrot16(%rip)
800	mov		\$10,%eax
801	movdqa		$xd0,0x100-0x100(%rcx)	# save SIMD counters
802	jmp		.Loop4x
803
804.align	32
805.Loop4x:
806___
807	foreach (&SSSE3_lane_ROUND(0, 4, 8,12)) { eval; }
808	foreach (&SSSE3_lane_ROUND(0, 5,10,15)) { eval; }
809$code.=<<___;
810	dec		%eax
811	jnz		.Loop4x
812
813	paddd		0x40(%rsp),$xa0		# accumulate key material
814	paddd		0x50(%rsp),$xa1
815	paddd		0x60(%rsp),$xa2
816	paddd		0x70(%rsp),$xa3
817
818	movdqa		$xa0,$xt2		# "de-interlace" data
819	punpckldq	$xa1,$xa0
820	movdqa		$xa2,$xt3
821	punpckldq	$xa3,$xa2
822	punpckhdq	$xa1,$xt2
823	punpckhdq	$xa3,$xt3
824	movdqa		$xa0,$xa1
825	punpcklqdq	$xa2,$xa0		# "a0"
826	movdqa		$xt2,$xa3
827	punpcklqdq	$xt3,$xt2		# "a2"
828	punpckhqdq	$xa2,$xa1		# "a1"
829	punpckhqdq	$xt3,$xa3		# "a3"
830___
831	($xa2,$xt2)=($xt2,$xa2);
832$code.=<<___;
833	paddd		0x80-0x100(%rcx),$xb0
834	paddd		0x90-0x100(%rcx),$xb1
835	paddd		0xa0-0x100(%rcx),$xb2
836	paddd		0xb0-0x100(%rcx),$xb3
837
838	movdqa		$xa0,0x00(%rsp)		# offload $xaN
839	movdqa		$xa1,0x10(%rsp)
840	movdqa		0x20(%rsp),$xa0		# "xc2"
841	movdqa		0x30(%rsp),$xa1		# "xc3"
842
843	movdqa		$xb0,$xt2
844	punpckldq	$xb1,$xb0
845	movdqa		$xb2,$xt3
846	punpckldq	$xb3,$xb2
847	punpckhdq	$xb1,$xt2
848	punpckhdq	$xb3,$xt3
849	movdqa		$xb0,$xb1
850	punpcklqdq	$xb2,$xb0		# "b0"
851	movdqa		$xt2,$xb3
852	punpcklqdq	$xt3,$xt2		# "b2"
853	punpckhqdq	$xb2,$xb1		# "b1"
854	punpckhqdq	$xt3,$xb3		# "b3"
855___
856	($xb2,$xt2)=($xt2,$xb2);
857	my ($xc0,$xc1,$xc2,$xc3)=($xt0,$xt1,$xa0,$xa1);
858$code.=<<___;
859	paddd		0xc0-0x100(%rcx),$xc0
860	paddd		0xd0-0x100(%rcx),$xc1
861	paddd		0xe0-0x100(%rcx),$xc2
862	paddd		0xf0-0x100(%rcx),$xc3
863
864	movdqa		$xa2,0x20(%rsp)		# keep offloading $xaN
865	movdqa		$xa3,0x30(%rsp)
866
867	movdqa		$xc0,$xt2
868	punpckldq	$xc1,$xc0
869	movdqa		$xc2,$xt3
870	punpckldq	$xc3,$xc2
871	punpckhdq	$xc1,$xt2
872	punpckhdq	$xc3,$xt3
873	movdqa		$xc0,$xc1
874	punpcklqdq	$xc2,$xc0		# "c0"
875	movdqa		$xt2,$xc3
876	punpcklqdq	$xt3,$xt2		# "c2"
877	punpckhqdq	$xc2,$xc1		# "c1"
878	punpckhqdq	$xt3,$xc3		# "c3"
879___
880	($xc2,$xt2)=($xt2,$xc2);
881	($xt0,$xt1)=($xa2,$xa3);		# use $xaN as temporary
882$code.=<<___;
883	paddd		0x100-0x100(%rcx),$xd0
884	paddd		0x110-0x100(%rcx),$xd1
885	paddd		0x120-0x100(%rcx),$xd2
886	paddd		0x130-0x100(%rcx),$xd3
887
888	movdqa		$xd0,$xt2
889	punpckldq	$xd1,$xd0
890	movdqa		$xd2,$xt3
891	punpckldq	$xd3,$xd2
892	punpckhdq	$xd1,$xt2
893	punpckhdq	$xd3,$xt3
894	movdqa		$xd0,$xd1
895	punpcklqdq	$xd2,$xd0		# "d0"
896	movdqa		$xt2,$xd3
897	punpcklqdq	$xt3,$xt2		# "d2"
898	punpckhqdq	$xd2,$xd1		# "d1"
899	punpckhqdq	$xt3,$xd3		# "d3"
900___
901	($xd2,$xt2)=($xt2,$xd2);
902$code.=<<___;
903	cmp		\$64*4,$len
904	jb		.Ltail4x
905
906	movdqu		0x00($inp),$xt0		# xor with input
907	movdqu		0x10($inp),$xt1
908	movdqu		0x20($inp),$xt2
909	movdqu		0x30($inp),$xt3
910	pxor		0x00(%rsp),$xt0		# $xaN is offloaded, remember?
911	pxor		$xb0,$xt1
912	pxor		$xc0,$xt2
913	pxor		$xd0,$xt3
914
915	 movdqu		$xt0,0x00($out)
916	movdqu		0x40($inp),$xt0
917	 movdqu		$xt1,0x10($out)
918	movdqu		0x50($inp),$xt1
919	 movdqu		$xt2,0x20($out)
920	movdqu		0x60($inp),$xt2
921	 movdqu		$xt3,0x30($out)
922	movdqu		0x70($inp),$xt3
923	lea		0x80($inp),$inp		# size optimization
924	pxor		0x10(%rsp),$xt0
925	pxor		$xb1,$xt1
926	pxor		$xc1,$xt2
927	pxor		$xd1,$xt3
928
929	 movdqu		$xt0,0x40($out)
930	movdqu		0x00($inp),$xt0
931	 movdqu		$xt1,0x50($out)
932	movdqu		0x10($inp),$xt1
933	 movdqu		$xt2,0x60($out)
934	movdqu		0x20($inp),$xt2
935	 movdqu		$xt3,0x70($out)
936	 lea		0x80($out),$out		# size optimization
937	movdqu		0x30($inp),$xt3
938	pxor		0x20(%rsp),$xt0
939	pxor		$xb2,$xt1
940	pxor		$xc2,$xt2
941	pxor		$xd2,$xt3
942
943	 movdqu		$xt0,0x00($out)
944	movdqu		0x40($inp),$xt0
945	 movdqu		$xt1,0x10($out)
946	movdqu		0x50($inp),$xt1
947	 movdqu		$xt2,0x20($out)
948	movdqu		0x60($inp),$xt2
949	 movdqu		$xt3,0x30($out)
950	movdqu		0x70($inp),$xt3
951	lea		0x80($inp),$inp		# inp+=64*4
952	pxor		0x30(%rsp),$xt0
953	pxor		$xb3,$xt1
954	pxor		$xc3,$xt2
955	pxor		$xd3,$xt3
956	movdqu		$xt0,0x40($out)
957	movdqu		$xt1,0x50($out)
958	movdqu		$xt2,0x60($out)
959	movdqu		$xt3,0x70($out)
960	lea		0x80($out),$out		# out+=64*4
961
962	sub		\$64*4,$len
963	jnz		.Loop_outer4x
964
965	jmp		.Ldone4x
966
967.Ltail4x:
968	cmp		\$192,$len
969	jae		.L192_or_more4x
970	cmp		\$128,$len
971	jae		.L128_or_more4x
972	cmp		\$64,$len
973	jae		.L64_or_more4x
974
975	#movdqa		0x00(%rsp),$xt0		# $xaN is offloaded, remember?
976	xor		%r10,%r10
977	#movdqa		$xt0,0x00(%rsp)
978	movdqa		$xb0,0x10(%rsp)
979	movdqa		$xc0,0x20(%rsp)
980	movdqa		$xd0,0x30(%rsp)
981	jmp		.Loop_tail4x
982
983.align	32
984.L64_or_more4x:
985	movdqu		0x00($inp),$xt0		# xor with input
986	movdqu		0x10($inp),$xt1
987	movdqu		0x20($inp),$xt2
988	movdqu		0x30($inp),$xt3
989	pxor		0x00(%rsp),$xt0		# $xaxN is offloaded, remember?
990	pxor		$xb0,$xt1
991	pxor		$xc0,$xt2
992	pxor		$xd0,$xt3
993	movdqu		$xt0,0x00($out)
994	movdqu		$xt1,0x10($out)
995	movdqu		$xt2,0x20($out)
996	movdqu		$xt3,0x30($out)
997	je		.Ldone4x
998
999	movdqa		0x10(%rsp),$xt0		# $xaN is offloaded, remember?
1000	lea		0x40($inp),$inp		# inp+=64*1
1001	xor		%r10,%r10
1002	movdqa		$xt0,0x00(%rsp)
1003	movdqa		$xb1,0x10(%rsp)
1004	lea		0x40($out),$out		# out+=64*1
1005	movdqa		$xc1,0x20(%rsp)
1006	sub		\$64,$len		# len-=64*1
1007	movdqa		$xd1,0x30(%rsp)
1008	jmp		.Loop_tail4x
1009
1010.align	32
1011.L128_or_more4x:
1012	movdqu		0x00($inp),$xt0		# xor with input
1013	movdqu		0x10($inp),$xt1
1014	movdqu		0x20($inp),$xt2
1015	movdqu		0x30($inp),$xt3
1016	pxor		0x00(%rsp),$xt0		# $xaN is offloaded, remember?
1017	pxor		$xb0,$xt1
1018	pxor		$xc0,$xt2
1019	pxor		$xd0,$xt3
1020
1021	 movdqu		$xt0,0x00($out)
1022	movdqu		0x40($inp),$xt0
1023	 movdqu		$xt1,0x10($out)
1024	movdqu		0x50($inp),$xt1
1025	 movdqu		$xt2,0x20($out)
1026	movdqu		0x60($inp),$xt2
1027	 movdqu		$xt3,0x30($out)
1028	movdqu		0x70($inp),$xt3
1029	pxor		0x10(%rsp),$xt0
1030	pxor		$xb1,$xt1
1031	pxor		$xc1,$xt2
1032	pxor		$xd1,$xt3
1033	movdqu		$xt0,0x40($out)
1034	movdqu		$xt1,0x50($out)
1035	movdqu		$xt2,0x60($out)
1036	movdqu		$xt3,0x70($out)
1037	je		.Ldone4x
1038
1039	movdqa		0x20(%rsp),$xt0		# $xaN is offloaded, remember?
1040	lea		0x80($inp),$inp		# inp+=64*2
1041	xor		%r10,%r10
1042	movdqa		$xt0,0x00(%rsp)
1043	movdqa		$xb2,0x10(%rsp)
1044	lea		0x80($out),$out		# out+=64*2
1045	movdqa		$xc2,0x20(%rsp)
1046	sub		\$128,$len		# len-=64*2
1047	movdqa		$xd2,0x30(%rsp)
1048	jmp		.Loop_tail4x
1049
1050.align	32
1051.L192_or_more4x:
1052	movdqu		0x00($inp),$xt0		# xor with input
1053	movdqu		0x10($inp),$xt1
1054	movdqu		0x20($inp),$xt2
1055	movdqu		0x30($inp),$xt3
1056	pxor		0x00(%rsp),$xt0		# $xaN is offloaded, remember?
1057	pxor		$xb0,$xt1
1058	pxor		$xc0,$xt2
1059	pxor		$xd0,$xt3
1060
1061	 movdqu		$xt0,0x00($out)
1062	movdqu		0x40($inp),$xt0
1063	 movdqu		$xt1,0x10($out)
1064	movdqu		0x50($inp),$xt1
1065	 movdqu		$xt2,0x20($out)
1066	movdqu		0x60($inp),$xt2
1067	 movdqu		$xt3,0x30($out)
1068	movdqu		0x70($inp),$xt3
1069	lea		0x80($inp),$inp		# size optimization
1070	pxor		0x10(%rsp),$xt0
1071	pxor		$xb1,$xt1
1072	pxor		$xc1,$xt2
1073	pxor		$xd1,$xt3
1074
1075	 movdqu		$xt0,0x40($out)
1076	movdqu		0x00($inp),$xt0
1077	 movdqu		$xt1,0x50($out)
1078	movdqu		0x10($inp),$xt1
1079	 movdqu		$xt2,0x60($out)
1080	movdqu		0x20($inp),$xt2
1081	 movdqu		$xt3,0x70($out)
1082	 lea		0x80($out),$out		# size optimization
1083	movdqu		0x30($inp),$xt3
1084	pxor		0x20(%rsp),$xt0
1085	pxor		$xb2,$xt1
1086	pxor		$xc2,$xt2
1087	pxor		$xd2,$xt3
1088	movdqu		$xt0,0x00($out)
1089	movdqu		$xt1,0x10($out)
1090	movdqu		$xt2,0x20($out)
1091	movdqu		$xt3,0x30($out)
1092	je		.Ldone4x
1093
1094	movdqa		0x30(%rsp),$xt0		# $xaN is offloaded, remember?
1095	lea		0x40($inp),$inp		# inp+=64*3
1096	xor		%r10,%r10
1097	movdqa		$xt0,0x00(%rsp)
1098	movdqa		$xb3,0x10(%rsp)
1099	lea		0x40($out),$out		# out+=64*3
1100	movdqa		$xc3,0x20(%rsp)
1101	sub		\$192,$len		# len-=64*3
1102	movdqa		$xd3,0x30(%rsp)
1103
1104.Loop_tail4x:
1105	movzb		($inp,%r10),%eax
1106	movzb		(%rsp,%r10),%ecx
1107	lea		1(%r10),%r10
1108	xor		%ecx,%eax
1109	mov		%al,-1($out,%r10)
1110	dec		$len
1111	jnz		.Loop_tail4x
1112
1113.Ldone4x:
1114___
1115$code.=<<___	if ($win64);
1116	movaps		-0xa8(%r9),%xmm6
1117	movaps		-0x98(%r9),%xmm7
1118	movaps		-0x88(%r9),%xmm8
1119	movaps		-0x78(%r9),%xmm9
1120	movaps		-0x68(%r9),%xmm10
1121	movaps		-0x58(%r9),%xmm11
1122	movaps		-0x48(%r9),%xmm12
1123	movaps		-0x38(%r9),%xmm13
1124	movaps		-0x28(%r9),%xmm14
1125	movaps		-0x18(%r9),%xmm15
1126___
1127$code.=<<___;
1128	lea		(%r9),%rsp
1129.cfi_def_cfa_register	rsp
1130.L4x_epilogue:
1131	ret
1132.cfi_endproc
1133.size	ChaCha20_ctr32_ssse3_4x,.-ChaCha20_ctr32_ssse3_4x
1134___
1135}
1136
1137########################################################################
1138# AVX2 code path
1139if ($avx>1) {
1140my ($xb0,$xb1,$xb2,$xb3, $xd0,$xd1,$xd2,$xd3,
1141    $xa0,$xa1,$xa2,$xa3, $xt0,$xt1,$xt2,$xt3)=map("%ymm$_",(0..15));
1142my @xx=($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
1143	"%nox","%nox","%nox","%nox", $xd0,$xd1,$xd2,$xd3);
1144
1145sub AVX2_lane_ROUND {
1146my ($a0,$b0,$c0,$d0)=@_;
1147my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
1148my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
1149my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
1150my ($xc,$xc_,$t0,$t1)=map("\"$_\"",$xt0,$xt1,$xt2,$xt3);
1151my @x=map("\"$_\"",@xx);
1152
1153	# Consider order in which variables are addressed by their
1154	# index:
1155	#
1156	#	a   b   c   d
1157	#
1158	#	0   4   8  12 < even round
1159	#	1   5   9  13
1160	#	2   6  10  14
1161	#	3   7  11  15
1162	#	0   5  10  15 < odd round
1163	#	1   6  11  12
1164	#	2   7   8  13
1165	#	3   4   9  14
1166	#
1167	# 'a', 'b' and 'd's are permanently allocated in registers,
1168	# @x[0..7,12..15], while 'c's are maintained in memory. If
1169	# you observe 'c' column, you'll notice that pair of 'c's is
1170	# invariant between rounds. This means that we have to reload
1171	# them once per round, in the middle. This is why you'll see
1172	# bunch of 'c' stores and loads in the middle, but none in
1173	# the beginning or end.
1174
1175	(
1176	"&vpaddd	(@x[$a0],@x[$a0],@x[$b0])",	# Q1
1177	"&vpxor		(@x[$d0],@x[$a0],@x[$d0])",
1178	"&vpshufb	(@x[$d0],@x[$d0],$t1)",
1179	 "&vpaddd	(@x[$a1],@x[$a1],@x[$b1])",	# Q2
1180	 "&vpxor	(@x[$d1],@x[$a1],@x[$d1])",
1181	 "&vpshufb	(@x[$d1],@x[$d1],$t1)",
1182
1183	"&vpaddd	($xc,$xc,@x[$d0])",
1184	"&vpxor		(@x[$b0],$xc,@x[$b0])",
1185	"&vpslld	($t0,@x[$b0],12)",
1186	"&vpsrld	(@x[$b0],@x[$b0],20)",
1187	"&vpor		(@x[$b0],$t0,@x[$b0])",
1188	"&vbroadcasti128($t0,'(%r11)')",		# .Lrot24(%rip)
1189	 "&vpaddd	($xc_,$xc_,@x[$d1])",
1190	 "&vpxor	(@x[$b1],$xc_,@x[$b1])",
1191	 "&vpslld	($t1,@x[$b1],12)",
1192	 "&vpsrld	(@x[$b1],@x[$b1],20)",
1193	 "&vpor		(@x[$b1],$t1,@x[$b1])",
1194
1195	"&vpaddd	(@x[$a0],@x[$a0],@x[$b0])",
1196	"&vpxor		(@x[$d0],@x[$a0],@x[$d0])",
1197	"&vpshufb	(@x[$d0],@x[$d0],$t0)",
1198	 "&vpaddd	(@x[$a1],@x[$a1],@x[$b1])",
1199	 "&vpxor	(@x[$d1],@x[$a1],@x[$d1])",
1200	 "&vpshufb	(@x[$d1],@x[$d1],$t0)",
1201
1202	"&vpaddd	($xc,$xc,@x[$d0])",
1203	"&vpxor		(@x[$b0],$xc,@x[$b0])",
1204	"&vpslld	($t1,@x[$b0],7)",
1205	"&vpsrld	(@x[$b0],@x[$b0],25)",
1206	"&vpor		(@x[$b0],$t1,@x[$b0])",
1207	"&vbroadcasti128($t1,'(%r10)')",		# .Lrot16(%rip)
1208	 "&vpaddd	($xc_,$xc_,@x[$d1])",
1209	 "&vpxor	(@x[$b1],$xc_,@x[$b1])",
1210	 "&vpslld	($t0,@x[$b1],7)",
1211	 "&vpsrld	(@x[$b1],@x[$b1],25)",
1212	 "&vpor		(@x[$b1],$t0,@x[$b1])",
1213
1214	"&vmovdqa	(\"`32*($c0-8)`(%rsp)\",$xc)",	# reload pair of 'c's
1215	 "&vmovdqa	(\"`32*($c1-8)`(%rsp)\",$xc_)",
1216	"&vmovdqa	($xc,\"`32*($c2-8)`(%rsp)\")",
1217	 "&vmovdqa	($xc_,\"`32*($c3-8)`(%rsp)\")",
1218
1219	"&vpaddd	(@x[$a2],@x[$a2],@x[$b2])",	# Q3
1220	"&vpxor		(@x[$d2],@x[$a2],@x[$d2])",
1221	"&vpshufb	(@x[$d2],@x[$d2],$t1)",
1222	 "&vpaddd	(@x[$a3],@x[$a3],@x[$b3])",	# Q4
1223	 "&vpxor	(@x[$d3],@x[$a3],@x[$d3])",
1224	 "&vpshufb	(@x[$d3],@x[$d3],$t1)",
1225
1226	"&vpaddd	($xc,$xc,@x[$d2])",
1227	"&vpxor		(@x[$b2],$xc,@x[$b2])",
1228	"&vpslld	($t0,@x[$b2],12)",
1229	"&vpsrld	(@x[$b2],@x[$b2],20)",
1230	"&vpor		(@x[$b2],$t0,@x[$b2])",
1231	"&vbroadcasti128($t0,'(%r11)')",		# .Lrot24(%rip)
1232	 "&vpaddd	($xc_,$xc_,@x[$d3])",
1233	 "&vpxor	(@x[$b3],$xc_,@x[$b3])",
1234	 "&vpslld	($t1,@x[$b3],12)",
1235	 "&vpsrld	(@x[$b3],@x[$b3],20)",
1236	 "&vpor		(@x[$b3],$t1,@x[$b3])",
1237
1238	"&vpaddd	(@x[$a2],@x[$a2],@x[$b2])",
1239	"&vpxor		(@x[$d2],@x[$a2],@x[$d2])",
1240	"&vpshufb	(@x[$d2],@x[$d2],$t0)",
1241	 "&vpaddd	(@x[$a3],@x[$a3],@x[$b3])",
1242	 "&vpxor	(@x[$d3],@x[$a3],@x[$d3])",
1243	 "&vpshufb	(@x[$d3],@x[$d3],$t0)",
1244
1245	"&vpaddd	($xc,$xc,@x[$d2])",
1246	"&vpxor		(@x[$b2],$xc,@x[$b2])",
1247	"&vpslld	($t1,@x[$b2],7)",
1248	"&vpsrld	(@x[$b2],@x[$b2],25)",
1249	"&vpor		(@x[$b2],$t1,@x[$b2])",
1250	"&vbroadcasti128($t1,'(%r10)')",		# .Lrot16(%rip)
1251	 "&vpaddd	($xc_,$xc_,@x[$d3])",
1252	 "&vpxor	(@x[$b3],$xc_,@x[$b3])",
1253	 "&vpslld	($t0,@x[$b3],7)",
1254	 "&vpsrld	(@x[$b3],@x[$b3],25)",
1255	 "&vpor		(@x[$b3],$t0,@x[$b3])"
1256	);
1257}
1258
1259my $xframe = $win64 ? 0xa8 : 8;
1260
1261$code.=<<___;
1262.globl	ChaCha20_ctr32_avx2
1263.type	ChaCha20_ctr32_avx2,\@function,5
1264.align	32
1265ChaCha20_ctr32_avx2:
1266.cfi_startproc
1267	_CET_ENDBR
1268	mov		%rsp,%r9		# frame register
1269.cfi_def_cfa_register	r9
1270	sub		\$0x280+$xframe,%rsp
1271	and		\$-32,%rsp
1272___
1273$code.=<<___	if ($win64);
1274	movaps		%xmm6,-0xa8(%r9)
1275	movaps		%xmm7,-0x98(%r9)
1276	movaps		%xmm8,-0x88(%r9)
1277	movaps		%xmm9,-0x78(%r9)
1278	movaps		%xmm10,-0x68(%r9)
1279	movaps		%xmm11,-0x58(%r9)
1280	movaps		%xmm12,-0x48(%r9)
1281	movaps		%xmm13,-0x38(%r9)
1282	movaps		%xmm14,-0x28(%r9)
1283	movaps		%xmm15,-0x18(%r9)
1284.L8x_body:
1285___
1286$code.=<<___;
1287	vzeroupper
1288
1289	################ stack layout
1290	# +0x00		SIMD equivalent of @x[8-12]
1291	# ...
1292	# +0x80		constant copy of key[0-2] smashed by lanes
1293	# ...
1294	# +0x200	SIMD counters (with nonce smashed by lanes)
1295	# ...
1296	# +0x280
1297
1298	vbroadcasti128	.Lsigma(%rip),$xa3	# key[0]
1299	vbroadcasti128	($key),$xb3		# key[1]
1300	vbroadcasti128	16($key),$xt3		# key[2]
1301	vbroadcasti128	($counter),$xd3		# key[3]
1302	lea		0x100(%rsp),%rcx	# size optimization
1303	lea		0x200(%rsp),%rax	# size optimization
1304	lea		.Lrot16(%rip),%r10
1305	lea		.Lrot24(%rip),%r11
1306
1307	vpshufd		\$0x00,$xa3,$xa0	# smash key by lanes...
1308	vpshufd		\$0x55,$xa3,$xa1
1309	vmovdqa		$xa0,0x80-0x100(%rcx)	# ... and offload
1310	vpshufd		\$0xaa,$xa3,$xa2
1311	vmovdqa		$xa1,0xa0-0x100(%rcx)
1312	vpshufd		\$0xff,$xa3,$xa3
1313	vmovdqa		$xa2,0xc0-0x100(%rcx)
1314	vmovdqa		$xa3,0xe0-0x100(%rcx)
1315
1316	vpshufd		\$0x00,$xb3,$xb0
1317	vpshufd		\$0x55,$xb3,$xb1
1318	vmovdqa		$xb0,0x100-0x100(%rcx)
1319	vpshufd		\$0xaa,$xb3,$xb2
1320	vmovdqa		$xb1,0x120-0x100(%rcx)
1321	vpshufd		\$0xff,$xb3,$xb3
1322	vmovdqa		$xb2,0x140-0x100(%rcx)
1323	vmovdqa		$xb3,0x160-0x100(%rcx)
1324
1325	vpshufd		\$0x00,$xt3,$xt0	# "xc0"
1326	vpshufd		\$0x55,$xt3,$xt1	# "xc1"
1327	vmovdqa		$xt0,0x180-0x200(%rax)
1328	vpshufd		\$0xaa,$xt3,$xt2	# "xc2"
1329	vmovdqa		$xt1,0x1a0-0x200(%rax)
1330	vpshufd		\$0xff,$xt3,$xt3	# "xc3"
1331	vmovdqa		$xt2,0x1c0-0x200(%rax)
1332	vmovdqa		$xt3,0x1e0-0x200(%rax)
1333
1334	vpshufd		\$0x00,$xd3,$xd0
1335	vpshufd		\$0x55,$xd3,$xd1
1336	vpaddd		.Lincy(%rip),$xd0,$xd0	# don't save counters yet
1337	vpshufd		\$0xaa,$xd3,$xd2
1338	vmovdqa		$xd1,0x220-0x200(%rax)
1339	vpshufd		\$0xff,$xd3,$xd3
1340	vmovdqa		$xd2,0x240-0x200(%rax)
1341	vmovdqa		$xd3,0x260-0x200(%rax)
1342
1343	jmp		.Loop_enter8x
1344
1345.align	32
1346.Loop_outer8x:
1347	vmovdqa		0x80-0x100(%rcx),$xa0	# re-load smashed key
1348	vmovdqa		0xa0-0x100(%rcx),$xa1
1349	vmovdqa		0xc0-0x100(%rcx),$xa2
1350	vmovdqa		0xe0-0x100(%rcx),$xa3
1351	vmovdqa		0x100-0x100(%rcx),$xb0
1352	vmovdqa		0x120-0x100(%rcx),$xb1
1353	vmovdqa		0x140-0x100(%rcx),$xb2
1354	vmovdqa		0x160-0x100(%rcx),$xb3
1355	vmovdqa		0x180-0x200(%rax),$xt0	# "xc0"
1356	vmovdqa		0x1a0-0x200(%rax),$xt1	# "xc1"
1357	vmovdqa		0x1c0-0x200(%rax),$xt2	# "xc2"
1358	vmovdqa		0x1e0-0x200(%rax),$xt3	# "xc3"
1359	vmovdqa		0x200-0x200(%rax),$xd0
1360	vmovdqa		0x220-0x200(%rax),$xd1
1361	vmovdqa		0x240-0x200(%rax),$xd2
1362	vmovdqa		0x260-0x200(%rax),$xd3
1363	vpaddd		.Leight(%rip),$xd0,$xd0	# next SIMD counters
1364
1365.Loop_enter8x:
1366	vmovdqa		$xt2,0x40(%rsp)		# SIMD equivalent of "@x[10]"
1367	vmovdqa		$xt3,0x60(%rsp)		# SIMD equivalent of "@x[11]"
1368	vbroadcasti128	(%r10),$xt3
1369	vmovdqa		$xd0,0x200-0x200(%rax)	# save SIMD counters
1370	mov		\$10,%eax
1371	jmp		.Loop8x
1372
1373.align	32
1374.Loop8x:
1375___
1376	foreach (&AVX2_lane_ROUND(0, 4, 8,12)) { eval; }
1377	foreach (&AVX2_lane_ROUND(0, 5,10,15)) { eval; }
1378$code.=<<___;
1379	dec		%eax
1380	jnz		.Loop8x
1381
1382	lea		0x200(%rsp),%rax	# size optimization
1383	vpaddd		0x80-0x100(%rcx),$xa0,$xa0	# accumulate key
1384	vpaddd		0xa0-0x100(%rcx),$xa1,$xa1
1385	vpaddd		0xc0-0x100(%rcx),$xa2,$xa2
1386	vpaddd		0xe0-0x100(%rcx),$xa3,$xa3
1387
1388	vpunpckldq	$xa1,$xa0,$xt2		# "de-interlace" data
1389	vpunpckldq	$xa3,$xa2,$xt3
1390	vpunpckhdq	$xa1,$xa0,$xa0
1391	vpunpckhdq	$xa3,$xa2,$xa2
1392	vpunpcklqdq	$xt3,$xt2,$xa1		# "a0"
1393	vpunpckhqdq	$xt3,$xt2,$xt2		# "a1"
1394	vpunpcklqdq	$xa2,$xa0,$xa3		# "a2"
1395	vpunpckhqdq	$xa2,$xa0,$xa0		# "a3"
1396___
1397	($xa0,$xa1,$xa2,$xa3,$xt2)=($xa1,$xt2,$xa3,$xa0,$xa2);
1398$code.=<<___;
1399	vpaddd		0x100-0x100(%rcx),$xb0,$xb0
1400	vpaddd		0x120-0x100(%rcx),$xb1,$xb1
1401	vpaddd		0x140-0x100(%rcx),$xb2,$xb2
1402	vpaddd		0x160-0x100(%rcx),$xb3,$xb3
1403
1404	vpunpckldq	$xb1,$xb0,$xt2
1405	vpunpckldq	$xb3,$xb2,$xt3
1406	vpunpckhdq	$xb1,$xb0,$xb0
1407	vpunpckhdq	$xb3,$xb2,$xb2
1408	vpunpcklqdq	$xt3,$xt2,$xb1		# "b0"
1409	vpunpckhqdq	$xt3,$xt2,$xt2		# "b1"
1410	vpunpcklqdq	$xb2,$xb0,$xb3		# "b2"
1411	vpunpckhqdq	$xb2,$xb0,$xb0		# "b3"
1412___
1413	($xb0,$xb1,$xb2,$xb3,$xt2)=($xb1,$xt2,$xb3,$xb0,$xb2);
1414$code.=<<___;
1415	vperm2i128	\$0x20,$xb0,$xa0,$xt3	# "de-interlace" further
1416	vperm2i128	\$0x31,$xb0,$xa0,$xb0
1417	vperm2i128	\$0x20,$xb1,$xa1,$xa0
1418	vperm2i128	\$0x31,$xb1,$xa1,$xb1
1419	vperm2i128	\$0x20,$xb2,$xa2,$xa1
1420	vperm2i128	\$0x31,$xb2,$xa2,$xb2
1421	vperm2i128	\$0x20,$xb3,$xa3,$xa2
1422	vperm2i128	\$0x31,$xb3,$xa3,$xb3
1423___
1424	($xa0,$xa1,$xa2,$xa3,$xt3)=($xt3,$xa0,$xa1,$xa2,$xa3);
1425	my ($xc0,$xc1,$xc2,$xc3)=($xt0,$xt1,$xa0,$xa1);
1426$code.=<<___;
1427	vmovdqa		$xa0,0x00(%rsp)		# offload $xaN
1428	vmovdqa		$xa1,0x20(%rsp)
1429	vmovdqa		0x40(%rsp),$xc2		# $xa0
1430	vmovdqa		0x60(%rsp),$xc3		# $xa1
1431
1432	vpaddd		0x180-0x200(%rax),$xc0,$xc0
1433	vpaddd		0x1a0-0x200(%rax),$xc1,$xc1
1434	vpaddd		0x1c0-0x200(%rax),$xc2,$xc2
1435	vpaddd		0x1e0-0x200(%rax),$xc3,$xc3
1436
1437	vpunpckldq	$xc1,$xc0,$xt2
1438	vpunpckldq	$xc3,$xc2,$xt3
1439	vpunpckhdq	$xc1,$xc0,$xc0
1440	vpunpckhdq	$xc3,$xc2,$xc2
1441	vpunpcklqdq	$xt3,$xt2,$xc1		# "c0"
1442	vpunpckhqdq	$xt3,$xt2,$xt2		# "c1"
1443	vpunpcklqdq	$xc2,$xc0,$xc3		# "c2"
1444	vpunpckhqdq	$xc2,$xc0,$xc0		# "c3"
1445___
1446	($xc0,$xc1,$xc2,$xc3,$xt2)=($xc1,$xt2,$xc3,$xc0,$xc2);
1447$code.=<<___;
1448	vpaddd		0x200-0x200(%rax),$xd0,$xd0
1449	vpaddd		0x220-0x200(%rax),$xd1,$xd1
1450	vpaddd		0x240-0x200(%rax),$xd2,$xd2
1451	vpaddd		0x260-0x200(%rax),$xd3,$xd3
1452
1453	vpunpckldq	$xd1,$xd0,$xt2
1454	vpunpckldq	$xd3,$xd2,$xt3
1455	vpunpckhdq	$xd1,$xd0,$xd0
1456	vpunpckhdq	$xd3,$xd2,$xd2
1457	vpunpcklqdq	$xt3,$xt2,$xd1		# "d0"
1458	vpunpckhqdq	$xt3,$xt2,$xt2		# "d1"
1459	vpunpcklqdq	$xd2,$xd0,$xd3		# "d2"
1460	vpunpckhqdq	$xd2,$xd0,$xd0		# "d3"
1461___
1462	($xd0,$xd1,$xd2,$xd3,$xt2)=($xd1,$xt2,$xd3,$xd0,$xd2);
1463$code.=<<___;
1464	vperm2i128	\$0x20,$xd0,$xc0,$xt3	# "de-interlace" further
1465	vperm2i128	\$0x31,$xd0,$xc0,$xd0
1466	vperm2i128	\$0x20,$xd1,$xc1,$xc0
1467	vperm2i128	\$0x31,$xd1,$xc1,$xd1
1468	vperm2i128	\$0x20,$xd2,$xc2,$xc1
1469	vperm2i128	\$0x31,$xd2,$xc2,$xd2
1470	vperm2i128	\$0x20,$xd3,$xc3,$xc2
1471	vperm2i128	\$0x31,$xd3,$xc3,$xd3
1472___
1473	($xc0,$xc1,$xc2,$xc3,$xt3)=($xt3,$xc0,$xc1,$xc2,$xc3);
1474	($xb0,$xb1,$xb2,$xb3,$xc0,$xc1,$xc2,$xc3)=
1475	($xc0,$xc1,$xc2,$xc3,$xb0,$xb1,$xb2,$xb3);
1476	($xa0,$xa1)=($xt2,$xt3);
1477$code.=<<___;
1478	vmovdqa		0x00(%rsp),$xa0		# $xaN was offloaded, remember?
1479	vmovdqa		0x20(%rsp),$xa1
1480
1481	cmp		\$64*8,$len
1482	jb		.Ltail8x
1483
1484	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1485	vpxor		0x20($inp),$xb0,$xb0
1486	vpxor		0x40($inp),$xc0,$xc0
1487	vpxor		0x60($inp),$xd0,$xd0
1488	lea		0x80($inp),$inp		# size optimization
1489	vmovdqu		$xa0,0x00($out)
1490	vmovdqu		$xb0,0x20($out)
1491	vmovdqu		$xc0,0x40($out)
1492	vmovdqu		$xd0,0x60($out)
1493	lea		0x80($out),$out		# size optimization
1494
1495	vpxor		0x00($inp),$xa1,$xa1
1496	vpxor		0x20($inp),$xb1,$xb1
1497	vpxor		0x40($inp),$xc1,$xc1
1498	vpxor		0x60($inp),$xd1,$xd1
1499	lea		0x80($inp),$inp		# size optimization
1500	vmovdqu		$xa1,0x00($out)
1501	vmovdqu		$xb1,0x20($out)
1502	vmovdqu		$xc1,0x40($out)
1503	vmovdqu		$xd1,0x60($out)
1504	lea		0x80($out),$out		# size optimization
1505
1506	vpxor		0x00($inp),$xa2,$xa2
1507	vpxor		0x20($inp),$xb2,$xb2
1508	vpxor		0x40($inp),$xc2,$xc2
1509	vpxor		0x60($inp),$xd2,$xd2
1510	lea		0x80($inp),$inp		# size optimization
1511	vmovdqu		$xa2,0x00($out)
1512	vmovdqu		$xb2,0x20($out)
1513	vmovdqu		$xc2,0x40($out)
1514	vmovdqu		$xd2,0x60($out)
1515	lea		0x80($out),$out		# size optimization
1516
1517	vpxor		0x00($inp),$xa3,$xa3
1518	vpxor		0x20($inp),$xb3,$xb3
1519	vpxor		0x40($inp),$xc3,$xc3
1520	vpxor		0x60($inp),$xd3,$xd3
1521	lea		0x80($inp),$inp		# size optimization
1522	vmovdqu		$xa3,0x00($out)
1523	vmovdqu		$xb3,0x20($out)
1524	vmovdqu		$xc3,0x40($out)
1525	vmovdqu		$xd3,0x60($out)
1526	lea		0x80($out),$out		# size optimization
1527
1528	sub		\$64*8,$len
1529	jnz		.Loop_outer8x
1530
1531	jmp		.Ldone8x
1532
1533.Ltail8x:
1534	cmp		\$448,$len
1535	jae		.L448_or_more8x
1536	cmp		\$384,$len
1537	jae		.L384_or_more8x
1538	cmp		\$320,$len
1539	jae		.L320_or_more8x
1540	cmp		\$256,$len
1541	jae		.L256_or_more8x
1542	cmp		\$192,$len
1543	jae		.L192_or_more8x
1544	cmp		\$128,$len
1545	jae		.L128_or_more8x
1546	cmp		\$64,$len
1547	jae		.L64_or_more8x
1548
1549	xor		%r10,%r10
1550	vmovdqa		$xa0,0x00(%rsp)
1551	vmovdqa		$xb0,0x20(%rsp)
1552	jmp		.Loop_tail8x
1553
1554.align	32
1555.L64_or_more8x:
1556	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1557	vpxor		0x20($inp),$xb0,$xb0
1558	vmovdqu		$xa0,0x00($out)
1559	vmovdqu		$xb0,0x20($out)
1560	je		.Ldone8x
1561
1562	lea		0x40($inp),$inp		# inp+=64*1
1563	xor		%r10,%r10
1564	vmovdqa		$xc0,0x00(%rsp)
1565	lea		0x40($out),$out		# out+=64*1
1566	sub		\$64,$len		# len-=64*1
1567	vmovdqa		$xd0,0x20(%rsp)
1568	jmp		.Loop_tail8x
1569
1570.align	32
1571.L128_or_more8x:
1572	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1573	vpxor		0x20($inp),$xb0,$xb0
1574	vpxor		0x40($inp),$xc0,$xc0
1575	vpxor		0x60($inp),$xd0,$xd0
1576	vmovdqu		$xa0,0x00($out)
1577	vmovdqu		$xb0,0x20($out)
1578	vmovdqu		$xc0,0x40($out)
1579	vmovdqu		$xd0,0x60($out)
1580	je		.Ldone8x
1581
1582	lea		0x80($inp),$inp		# inp+=64*2
1583	xor		%r10,%r10
1584	vmovdqa		$xa1,0x00(%rsp)
1585	lea		0x80($out),$out		# out+=64*2
1586	sub		\$128,$len		# len-=64*2
1587	vmovdqa		$xb1,0x20(%rsp)
1588	jmp		.Loop_tail8x
1589
1590.align	32
1591.L192_or_more8x:
1592	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1593	vpxor		0x20($inp),$xb0,$xb0
1594	vpxor		0x40($inp),$xc0,$xc0
1595	vpxor		0x60($inp),$xd0,$xd0
1596	vpxor		0x80($inp),$xa1,$xa1
1597	vpxor		0xa0($inp),$xb1,$xb1
1598	vmovdqu		$xa0,0x00($out)
1599	vmovdqu		$xb0,0x20($out)
1600	vmovdqu		$xc0,0x40($out)
1601	vmovdqu		$xd0,0x60($out)
1602	vmovdqu		$xa1,0x80($out)
1603	vmovdqu		$xb1,0xa0($out)
1604	je		.Ldone8x
1605
1606	lea		0xc0($inp),$inp		# inp+=64*3
1607	xor		%r10,%r10
1608	vmovdqa		$xc1,0x00(%rsp)
1609	lea		0xc0($out),$out		# out+=64*3
1610	sub		\$192,$len		# len-=64*3
1611	vmovdqa		$xd1,0x20(%rsp)
1612	jmp		.Loop_tail8x
1613
1614.align	32
1615.L256_or_more8x:
1616	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1617	vpxor		0x20($inp),$xb0,$xb0
1618	vpxor		0x40($inp),$xc0,$xc0
1619	vpxor		0x60($inp),$xd0,$xd0
1620	vpxor		0x80($inp),$xa1,$xa1
1621	vpxor		0xa0($inp),$xb1,$xb1
1622	vpxor		0xc0($inp),$xc1,$xc1
1623	vpxor		0xe0($inp),$xd1,$xd1
1624	vmovdqu		$xa0,0x00($out)
1625	vmovdqu		$xb0,0x20($out)
1626	vmovdqu		$xc0,0x40($out)
1627	vmovdqu		$xd0,0x60($out)
1628	vmovdqu		$xa1,0x80($out)
1629	vmovdqu		$xb1,0xa0($out)
1630	vmovdqu		$xc1,0xc0($out)
1631	vmovdqu		$xd1,0xe0($out)
1632	je		.Ldone8x
1633
1634	lea		0x100($inp),$inp	# inp+=64*4
1635	xor		%r10,%r10
1636	vmovdqa		$xa2,0x00(%rsp)
1637	lea		0x100($out),$out	# out+=64*4
1638	sub		\$256,$len		# len-=64*4
1639	vmovdqa		$xb2,0x20(%rsp)
1640	jmp		.Loop_tail8x
1641
1642.align	32
1643.L320_or_more8x:
1644	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1645	vpxor		0x20($inp),$xb0,$xb0
1646	vpxor		0x40($inp),$xc0,$xc0
1647	vpxor		0x60($inp),$xd0,$xd0
1648	vpxor		0x80($inp),$xa1,$xa1
1649	vpxor		0xa0($inp),$xb1,$xb1
1650	vpxor		0xc0($inp),$xc1,$xc1
1651	vpxor		0xe0($inp),$xd1,$xd1
1652	vpxor		0x100($inp),$xa2,$xa2
1653	vpxor		0x120($inp),$xb2,$xb2
1654	vmovdqu		$xa0,0x00($out)
1655	vmovdqu		$xb0,0x20($out)
1656	vmovdqu		$xc0,0x40($out)
1657	vmovdqu		$xd0,0x60($out)
1658	vmovdqu		$xa1,0x80($out)
1659	vmovdqu		$xb1,0xa0($out)
1660	vmovdqu		$xc1,0xc0($out)
1661	vmovdqu		$xd1,0xe0($out)
1662	vmovdqu		$xa2,0x100($out)
1663	vmovdqu		$xb2,0x120($out)
1664	je		.Ldone8x
1665
1666	lea		0x140($inp),$inp	# inp+=64*5
1667	xor		%r10,%r10
1668	vmovdqa		$xc2,0x00(%rsp)
1669	lea		0x140($out),$out	# out+=64*5
1670	sub		\$320,$len		# len-=64*5
1671	vmovdqa		$xd2,0x20(%rsp)
1672	jmp		.Loop_tail8x
1673
1674.align	32
1675.L384_or_more8x:
1676	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1677	vpxor		0x20($inp),$xb0,$xb0
1678	vpxor		0x40($inp),$xc0,$xc0
1679	vpxor		0x60($inp),$xd0,$xd0
1680	vpxor		0x80($inp),$xa1,$xa1
1681	vpxor		0xa0($inp),$xb1,$xb1
1682	vpxor		0xc0($inp),$xc1,$xc1
1683	vpxor		0xe0($inp),$xd1,$xd1
1684	vpxor		0x100($inp),$xa2,$xa2
1685	vpxor		0x120($inp),$xb2,$xb2
1686	vpxor		0x140($inp),$xc2,$xc2
1687	vpxor		0x160($inp),$xd2,$xd2
1688	vmovdqu		$xa0,0x00($out)
1689	vmovdqu		$xb0,0x20($out)
1690	vmovdqu		$xc0,0x40($out)
1691	vmovdqu		$xd0,0x60($out)
1692	vmovdqu		$xa1,0x80($out)
1693	vmovdqu		$xb1,0xa0($out)
1694	vmovdqu		$xc1,0xc0($out)
1695	vmovdqu		$xd1,0xe0($out)
1696	vmovdqu		$xa2,0x100($out)
1697	vmovdqu		$xb2,0x120($out)
1698	vmovdqu		$xc2,0x140($out)
1699	vmovdqu		$xd2,0x160($out)
1700	je		.Ldone8x
1701
1702	lea		0x180($inp),$inp	# inp+=64*6
1703	xor		%r10,%r10
1704	vmovdqa		$xa3,0x00(%rsp)
1705	lea		0x180($out),$out	# out+=64*6
1706	sub		\$384,$len		# len-=64*6
1707	vmovdqa		$xb3,0x20(%rsp)
1708	jmp		.Loop_tail8x
1709
1710.align	32
1711.L448_or_more8x:
1712	vpxor		0x00($inp),$xa0,$xa0	# xor with input
1713	vpxor		0x20($inp),$xb0,$xb0
1714	vpxor		0x40($inp),$xc0,$xc0
1715	vpxor		0x60($inp),$xd0,$xd0
1716	vpxor		0x80($inp),$xa1,$xa1
1717	vpxor		0xa0($inp),$xb1,$xb1
1718	vpxor		0xc0($inp),$xc1,$xc1
1719	vpxor		0xe0($inp),$xd1,$xd1
1720	vpxor		0x100($inp),$xa2,$xa2
1721	vpxor		0x120($inp),$xb2,$xb2
1722	vpxor		0x140($inp),$xc2,$xc2
1723	vpxor		0x160($inp),$xd2,$xd2
1724	vpxor		0x180($inp),$xa3,$xa3
1725	vpxor		0x1a0($inp),$xb3,$xb3
1726	vmovdqu		$xa0,0x00($out)
1727	vmovdqu		$xb0,0x20($out)
1728	vmovdqu		$xc0,0x40($out)
1729	vmovdqu		$xd0,0x60($out)
1730	vmovdqu		$xa1,0x80($out)
1731	vmovdqu		$xb1,0xa0($out)
1732	vmovdqu		$xc1,0xc0($out)
1733	vmovdqu		$xd1,0xe0($out)
1734	vmovdqu		$xa2,0x100($out)
1735	vmovdqu		$xb2,0x120($out)
1736	vmovdqu		$xc2,0x140($out)
1737	vmovdqu		$xd2,0x160($out)
1738	vmovdqu		$xa3,0x180($out)
1739	vmovdqu		$xb3,0x1a0($out)
1740	je		.Ldone8x
1741
1742	lea		0x1c0($inp),$inp	# inp+=64*7
1743	xor		%r10,%r10
1744	vmovdqa		$xc3,0x00(%rsp)
1745	lea		0x1c0($out),$out	# out+=64*7
1746	sub		\$448,$len		# len-=64*7
1747	vmovdqa		$xd3,0x20(%rsp)
1748
1749.Loop_tail8x:
1750	movzb		($inp,%r10),%eax
1751	movzb		(%rsp,%r10),%ecx
1752	lea		1(%r10),%r10
1753	xor		%ecx,%eax
1754	mov		%al,-1($out,%r10)
1755	dec		$len
1756	jnz		.Loop_tail8x
1757
1758.Ldone8x:
1759	vzeroall
1760___
1761$code.=<<___	if ($win64);
1762	movaps		-0xa8(%r9),%xmm6
1763	movaps		-0x98(%r9),%xmm7
1764	movaps		-0x88(%r9),%xmm8
1765	movaps		-0x78(%r9),%xmm9
1766	movaps		-0x68(%r9),%xmm10
1767	movaps		-0x58(%r9),%xmm11
1768	movaps		-0x48(%r9),%xmm12
1769	movaps		-0x38(%r9),%xmm13
1770	movaps		-0x28(%r9),%xmm14
1771	movaps		-0x18(%r9),%xmm15
1772___
1773$code.=<<___;
1774	lea		(%r9),%rsp
1775.cfi_def_cfa_register	rsp
1776.L8x_epilogue:
1777	ret
1778.cfi_endproc
1779.size	ChaCha20_ctr32_avx2,.-ChaCha20_ctr32_avx2
1780___
1781}
1782
1783########################################################################
1784# AVX512 code paths
1785if ($avx>2) {
1786# This one handles shorter inputs...
1787
1788my ($a,$b,$c,$d, $a_,$b_,$c_,$d_,$fourz) = map("%zmm$_",(0..3,16..20));
1789my ($t0,$t1,$t2,$t3) = map("%xmm$_",(4..7));
1790
1791sub AVX512ROUND {	# critical path is 14 "SIMD ticks" per round
1792	&vpaddd	($a,$a,$b);
1793	&vpxord	($d,$d,$a);
1794	&vprold	($d,$d,16);
1795
1796	&vpaddd	($c,$c,$d);
1797	&vpxord	($b,$b,$c);
1798	&vprold	($b,$b,12);
1799
1800	&vpaddd	($a,$a,$b);
1801	&vpxord	($d,$d,$a);
1802	&vprold	($d,$d,8);
1803
1804	&vpaddd	($c,$c,$d);
1805	&vpxord	($b,$b,$c);
1806	&vprold	($b,$b,7);
1807}
1808
1809my $xframe = $win64 ? 32+8 : 8;
1810
1811$code.=<<___;
1812.type	ChaCha20_avx512,\@function,5
1813.align	32
1814ChaCha20_avx512:
1815.LChaCha20_avx512:
1816.cfi_startproc
1817	mov	%rsp,%r9		# frame pointer
1818.cfi_def_cfa_register	r9
1819	cmp	\$512,$len
1820	ja	.LChaCha20_16x
1821
1822	sub	\$64+$xframe,%rsp
1823___
1824$code.=<<___	if ($win64);
1825	movaps	%xmm6,-0x28(%r9)
1826	movaps	%xmm7,-0x18(%r9)
1827.Lavx512_body:
1828___
1829$code.=<<___;
1830	vbroadcasti32x4	.Lsigma(%rip),$a
1831	vbroadcasti32x4	($key),$b
1832	vbroadcasti32x4	16($key),$c
1833	vbroadcasti32x4	($counter),$d
1834
1835	vmovdqa32	$a,$a_
1836	vmovdqa32	$b,$b_
1837	vmovdqa32	$c,$c_
1838	vpaddd		.Lzeroz(%rip),$d,$d
1839	vmovdqa32	.Lfourz(%rip),$fourz
1840	mov		\$10,$counter	# reuse $counter
1841	vmovdqa32	$d,$d_
1842	jmp		.Loop_avx512
1843
1844.align	16
1845.Loop_outer_avx512:
1846	vmovdqa32	$a_,$a
1847	vmovdqa32	$b_,$b
1848	vmovdqa32	$c_,$c
1849	vpaddd		$fourz,$d_,$d
1850	mov		\$10,$counter
1851	vmovdqa32	$d,$d_
1852	jmp		.Loop_avx512
1853
1854.align	32
1855.Loop_avx512:
1856___
1857	&AVX512ROUND();
1858	&vpshufd	($c,$c,0b01001110);
1859	&vpshufd	($b,$b,0b00111001);
1860	&vpshufd	($d,$d,0b10010011);
1861
1862	&AVX512ROUND();
1863	&vpshufd	($c,$c,0b01001110);
1864	&vpshufd	($b,$b,0b10010011);
1865	&vpshufd	($d,$d,0b00111001);
1866
1867	&dec		($counter);
1868	&jnz		(".Loop_avx512");
1869
1870$code.=<<___;
1871	vpaddd		$a_,$a,$a
1872	vpaddd		$b_,$b,$b
1873	vpaddd		$c_,$c,$c
1874	vpaddd		$d_,$d,$d
1875
1876	sub		\$64,$len
1877	jb		.Ltail64_avx512
1878
1879	vpxor		0x00($inp),%x#$a,$t0	# xor with input
1880	vpxor		0x10($inp),%x#$b,$t1
1881	vpxor		0x20($inp),%x#$c,$t2
1882	vpxor		0x30($inp),%x#$d,$t3
1883	lea		0x40($inp),$inp		# inp+=64
1884
1885	vmovdqu		$t0,0x00($out)		# write output
1886	vmovdqu		$t1,0x10($out)
1887	vmovdqu		$t2,0x20($out)
1888	vmovdqu		$t3,0x30($out)
1889	lea		0x40($out),$out		# out+=64
1890
1891	jz		.Ldone_avx512
1892
1893	vextracti32x4	\$1,$a,$t0
1894	vextracti32x4	\$1,$b,$t1
1895	vextracti32x4	\$1,$c,$t2
1896	vextracti32x4	\$1,$d,$t3
1897
1898	sub		\$64,$len
1899	jb		.Ltail_avx512
1900
1901	vpxor		0x00($inp),$t0,$t0	# xor with input
1902	vpxor		0x10($inp),$t1,$t1
1903	vpxor		0x20($inp),$t2,$t2
1904	vpxor		0x30($inp),$t3,$t3
1905	lea		0x40($inp),$inp		# inp+=64
1906
1907	vmovdqu		$t0,0x00($out)		# write output
1908	vmovdqu		$t1,0x10($out)
1909	vmovdqu		$t2,0x20($out)
1910	vmovdqu		$t3,0x30($out)
1911	lea		0x40($out),$out		# out+=64
1912
1913	jz		.Ldone_avx512
1914
1915	vextracti32x4	\$2,$a,$t0
1916	vextracti32x4	\$2,$b,$t1
1917	vextracti32x4	\$2,$c,$t2
1918	vextracti32x4	\$2,$d,$t3
1919
1920	sub		\$64,$len
1921	jb		.Ltail_avx512
1922
1923	vpxor		0x00($inp),$t0,$t0	# xor with input
1924	vpxor		0x10($inp),$t1,$t1
1925	vpxor		0x20($inp),$t2,$t2
1926	vpxor		0x30($inp),$t3,$t3
1927	lea		0x40($inp),$inp		# inp+=64
1928
1929	vmovdqu		$t0,0x00($out)		# write output
1930	vmovdqu		$t1,0x10($out)
1931	vmovdqu		$t2,0x20($out)
1932	vmovdqu		$t3,0x30($out)
1933	lea		0x40($out),$out		# out+=64
1934
1935	jz		.Ldone_avx512
1936
1937	vextracti32x4	\$3,$a,$t0
1938	vextracti32x4	\$3,$b,$t1
1939	vextracti32x4	\$3,$c,$t2
1940	vextracti32x4	\$3,$d,$t3
1941
1942	sub		\$64,$len
1943	jb		.Ltail_avx512
1944
1945	vpxor		0x00($inp),$t0,$t0	# xor with input
1946	vpxor		0x10($inp),$t1,$t1
1947	vpxor		0x20($inp),$t2,$t2
1948	vpxor		0x30($inp),$t3,$t3
1949	lea		0x40($inp),$inp		# inp+=64
1950
1951	vmovdqu		$t0,0x00($out)		# write output
1952	vmovdqu		$t1,0x10($out)
1953	vmovdqu		$t2,0x20($out)
1954	vmovdqu		$t3,0x30($out)
1955	lea		0x40($out),$out		# out+=64
1956
1957	jnz		.Loop_outer_avx512
1958
1959	jmp		.Ldone_avx512
1960
1961.align	16
1962.Ltail64_avx512:
1963	vmovdqa		%x#$a,0x00(%rsp)
1964	vmovdqa		%x#$b,0x10(%rsp)
1965	vmovdqa		%x#$c,0x20(%rsp)
1966	vmovdqa		%x#$d,0x30(%rsp)
1967	add		\$64,$len
1968	jmp		.Loop_tail_avx512
1969
1970.align	16
1971.Ltail_avx512:
1972	vmovdqa		$t0,0x00(%rsp)
1973	vmovdqa		$t1,0x10(%rsp)
1974	vmovdqa		$t2,0x20(%rsp)
1975	vmovdqa		$t3,0x30(%rsp)
1976	add		\$64,$len
1977
1978.Loop_tail_avx512:
1979	movzb		($inp,$counter),%eax
1980	movzb		(%rsp,$counter),%ecx
1981	lea		1($counter),$counter
1982	xor		%ecx,%eax
1983	mov		%al,-1($out,$counter)
1984	dec		$len
1985	jnz		.Loop_tail_avx512
1986
1987	vmovdqa32	$a_,0x00(%rsp)
1988
1989.Ldone_avx512:
1990	vzeroall
1991___
1992$code.=<<___	if ($win64);
1993	movaps	-0x28(%r9),%xmm6
1994	movaps	-0x18(%r9),%xmm7
1995___
1996$code.=<<___;
1997	lea	(%r9),%rsp
1998.cfi_def_cfa_register	rsp
1999.Lavx512_epilogue:
2000	ret
2001.cfi_endproc
2002.size	ChaCha20_avx512,.-ChaCha20_avx512
2003___
2004}
2005if ($avx>2) {
2006# This one handles longer inputs...
2007
2008my ($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
2009    $xc0,$xc1,$xc2,$xc3, $xd0,$xd1,$xd2,$xd3)=map("%zmm$_",(0..15));
2010my  @xx=($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
2011	 $xc0,$xc1,$xc2,$xc3, $xd0,$xd1,$xd2,$xd3);
2012my @key=map("%zmm$_",(16..31));
2013my ($xt0,$xt1,$xt2,$xt3)=@key[0..3];
2014
2015sub AVX512_lane_ROUND {
2016my ($a0,$b0,$c0,$d0)=@_;
2017my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0));
2018my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1));
2019my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2));
2020my @x=map("\"$_\"",@xx);
2021
2022	(
2023	"&vpaddd	(@x[$a0],@x[$a0],@x[$b0])",	# Q1
2024	 "&vpaddd	(@x[$a1],@x[$a1],@x[$b1])",	# Q2
2025	  "&vpaddd	(@x[$a2],@x[$a2],@x[$b2])",	# Q3
2026	   "&vpaddd	(@x[$a3],@x[$a3],@x[$b3])",	# Q4
2027	"&vpxord	(@x[$d0],@x[$d0],@x[$a0])",
2028	 "&vpxord	(@x[$d1],@x[$d1],@x[$a1])",
2029	  "&vpxord	(@x[$d2],@x[$d2],@x[$a2])",
2030	   "&vpxord	(@x[$d3],@x[$d3],@x[$a3])",
2031	"&vprold	(@x[$d0],@x[$d0],16)",
2032	 "&vprold	(@x[$d1],@x[$d1],16)",
2033	  "&vprold	(@x[$d2],@x[$d2],16)",
2034	   "&vprold	(@x[$d3],@x[$d3],16)",
2035
2036	"&vpaddd	(@x[$c0],@x[$c0],@x[$d0])",
2037	 "&vpaddd	(@x[$c1],@x[$c1],@x[$d1])",
2038	  "&vpaddd	(@x[$c2],@x[$c2],@x[$d2])",
2039	   "&vpaddd	(@x[$c3],@x[$c3],@x[$d3])",
2040	"&vpxord	(@x[$b0],@x[$b0],@x[$c0])",
2041	 "&vpxord	(@x[$b1],@x[$b1],@x[$c1])",
2042	  "&vpxord	(@x[$b2],@x[$b2],@x[$c2])",
2043	   "&vpxord	(@x[$b3],@x[$b3],@x[$c3])",
2044	"&vprold	(@x[$b0],@x[$b0],12)",
2045	 "&vprold	(@x[$b1],@x[$b1],12)",
2046	  "&vprold	(@x[$b2],@x[$b2],12)",
2047	   "&vprold	(@x[$b3],@x[$b3],12)",
2048
2049	"&vpaddd	(@x[$a0],@x[$a0],@x[$b0])",
2050	 "&vpaddd	(@x[$a1],@x[$a1],@x[$b1])",
2051	  "&vpaddd	(@x[$a2],@x[$a2],@x[$b2])",
2052	   "&vpaddd	(@x[$a3],@x[$a3],@x[$b3])",
2053	"&vpxord	(@x[$d0],@x[$d0],@x[$a0])",
2054	 "&vpxord	(@x[$d1],@x[$d1],@x[$a1])",
2055	  "&vpxord	(@x[$d2],@x[$d2],@x[$a2])",
2056	   "&vpxord	(@x[$d3],@x[$d3],@x[$a3])",
2057	"&vprold	(@x[$d0],@x[$d0],8)",
2058	 "&vprold	(@x[$d1],@x[$d1],8)",
2059	  "&vprold	(@x[$d2],@x[$d2],8)",
2060	   "&vprold	(@x[$d3],@x[$d3],8)",
2061
2062	"&vpaddd	(@x[$c0],@x[$c0],@x[$d0])",
2063	 "&vpaddd	(@x[$c1],@x[$c1],@x[$d1])",
2064	  "&vpaddd	(@x[$c2],@x[$c2],@x[$d2])",
2065	   "&vpaddd	(@x[$c3],@x[$c3],@x[$d3])",
2066	"&vpxord	(@x[$b0],@x[$b0],@x[$c0])",
2067	 "&vpxord	(@x[$b1],@x[$b1],@x[$c1])",
2068	  "&vpxord	(@x[$b2],@x[$b2],@x[$c2])",
2069	   "&vpxord	(@x[$b3],@x[$b3],@x[$c3])",
2070	"&vprold	(@x[$b0],@x[$b0],7)",
2071	 "&vprold	(@x[$b1],@x[$b1],7)",
2072	  "&vprold	(@x[$b2],@x[$b2],7)",
2073	   "&vprold	(@x[$b3],@x[$b3],7)"
2074	);
2075}
2076
2077my $xframe = $win64 ? 0xa8 : 8;
2078
2079$code.=<<___;
2080.type	ChaCha20_16x,\@function,5
2081.align	32
2082ChaCha20_16x:
2083.LChaCha20_16x:
2084.cfi_startproc
2085	mov		%rsp,%r9		# frame register
2086.cfi_def_cfa_register	r9
2087	sub		\$64+$xframe,%rsp
2088	and		\$-64,%rsp
2089___
2090$code.=<<___	if ($win64);
2091	movaps		%xmm6,-0xa8(%r9)
2092	movaps		%xmm7,-0x98(%r9)
2093	movaps		%xmm8,-0x88(%r9)
2094	movaps		%xmm9,-0x78(%r9)
2095	movaps		%xmm10,-0x68(%r9)
2096	movaps		%xmm11,-0x58(%r9)
2097	movaps		%xmm12,-0x48(%r9)
2098	movaps		%xmm13,-0x38(%r9)
2099	movaps		%xmm14,-0x28(%r9)
2100	movaps		%xmm15,-0x18(%r9)
2101.L16x_body:
2102___
2103$code.=<<___;
2104	vzeroupper
2105
2106	lea		.Lsigma(%rip),%r10
2107	vbroadcasti32x4	(%r10),$xa3		# key[0]
2108	vbroadcasti32x4	($key),$xb3		# key[1]
2109	vbroadcasti32x4	16($key),$xc3		# key[2]
2110	vbroadcasti32x4	($counter),$xd3		# key[3]
2111
2112	vpshufd		\$0x00,$xa3,$xa0	# smash key by lanes...
2113	vpshufd		\$0x55,$xa3,$xa1
2114	vpshufd		\$0xaa,$xa3,$xa2
2115	vpshufd		\$0xff,$xa3,$xa3
2116	vmovdqa64	$xa0,@key[0]
2117	vmovdqa64	$xa1,@key[1]
2118	vmovdqa64	$xa2,@key[2]
2119	vmovdqa64	$xa3,@key[3]
2120
2121	vpshufd		\$0x00,$xb3,$xb0
2122	vpshufd		\$0x55,$xb3,$xb1
2123	vpshufd		\$0xaa,$xb3,$xb2
2124	vpshufd		\$0xff,$xb3,$xb3
2125	vmovdqa64	$xb0,@key[4]
2126	vmovdqa64	$xb1,@key[5]
2127	vmovdqa64	$xb2,@key[6]
2128	vmovdqa64	$xb3,@key[7]
2129
2130	vpshufd		\$0x00,$xc3,$xc0
2131	vpshufd		\$0x55,$xc3,$xc1
2132	vpshufd		\$0xaa,$xc3,$xc2
2133	vpshufd		\$0xff,$xc3,$xc3
2134	vmovdqa64	$xc0,@key[8]
2135	vmovdqa64	$xc1,@key[9]
2136	vmovdqa64	$xc2,@key[10]
2137	vmovdqa64	$xc3,@key[11]
2138
2139	vpshufd		\$0x00,$xd3,$xd0
2140	vpshufd		\$0x55,$xd3,$xd1
2141	vpshufd		\$0xaa,$xd3,$xd2
2142	vpshufd		\$0xff,$xd3,$xd3
2143	vpaddd		.Lincz(%rip),$xd0,$xd0	# don't save counters yet
2144	vmovdqa64	$xd0,@key[12]
2145	vmovdqa64	$xd1,@key[13]
2146	vmovdqa64	$xd2,@key[14]
2147	vmovdqa64	$xd3,@key[15]
2148
2149	mov		\$10,%eax
2150	jmp		.Loop16x
2151
2152.align	32
2153.Loop_outer16x:
2154	vpbroadcastd	0(%r10),$xa0		# reload key
2155	vpbroadcastd	4(%r10),$xa1
2156	vpbroadcastd	8(%r10),$xa2
2157	vpbroadcastd	12(%r10),$xa3
2158	vpaddd		.Lsixteen(%rip),@key[12],@key[12]	# next SIMD counters
2159	vmovdqa64	@key[4],$xb0
2160	vmovdqa64	@key[5],$xb1
2161	vmovdqa64	@key[6],$xb2
2162	vmovdqa64	@key[7],$xb3
2163	vmovdqa64	@key[8],$xc0
2164	vmovdqa64	@key[9],$xc1
2165	vmovdqa64	@key[10],$xc2
2166	vmovdqa64	@key[11],$xc3
2167	vmovdqa64	@key[12],$xd0
2168	vmovdqa64	@key[13],$xd1
2169	vmovdqa64	@key[14],$xd2
2170	vmovdqa64	@key[15],$xd3
2171
2172	vmovdqa64	$xa0,@key[0]
2173	vmovdqa64	$xa1,@key[1]
2174	vmovdqa64	$xa2,@key[2]
2175	vmovdqa64	$xa3,@key[3]
2176
2177	mov		\$10,%eax
2178	jmp		.Loop16x
2179
2180.align	32
2181.Loop16x:
2182___
2183	foreach (&AVX512_lane_ROUND(0, 4, 8,12)) { eval; }
2184	foreach (&AVX512_lane_ROUND(0, 5,10,15)) { eval; }
2185$code.=<<___;
2186	dec		%eax
2187	jnz		.Loop16x
2188
2189	vpaddd		@key[0],$xa0,$xa0	# accumulate key
2190	vpaddd		@key[1],$xa1,$xa1
2191	vpaddd		@key[2],$xa2,$xa2
2192	vpaddd		@key[3],$xa3,$xa3
2193
2194	vpunpckldq	$xa1,$xa0,$xt2		# "de-interlace" data
2195	vpunpckldq	$xa3,$xa2,$xt3
2196	vpunpckhdq	$xa1,$xa0,$xa0
2197	vpunpckhdq	$xa3,$xa2,$xa2
2198	vpunpcklqdq	$xt3,$xt2,$xa1		# "a0"
2199	vpunpckhqdq	$xt3,$xt2,$xt2		# "a1"
2200	vpunpcklqdq	$xa2,$xa0,$xa3		# "a2"
2201	vpunpckhqdq	$xa2,$xa0,$xa0		# "a3"
2202___
2203	($xa0,$xa1,$xa2,$xa3,$xt2)=($xa1,$xt2,$xa3,$xa0,$xa2);
2204$code.=<<___;
2205	vpaddd		@key[4],$xb0,$xb0
2206	vpaddd		@key[5],$xb1,$xb1
2207	vpaddd		@key[6],$xb2,$xb2
2208	vpaddd		@key[7],$xb3,$xb3
2209
2210	vpunpckldq	$xb1,$xb0,$xt2
2211	vpunpckldq	$xb3,$xb2,$xt3
2212	vpunpckhdq	$xb1,$xb0,$xb0
2213	vpunpckhdq	$xb3,$xb2,$xb2
2214	vpunpcklqdq	$xt3,$xt2,$xb1		# "b0"
2215	vpunpckhqdq	$xt3,$xt2,$xt2		# "b1"
2216	vpunpcklqdq	$xb2,$xb0,$xb3		# "b2"
2217	vpunpckhqdq	$xb2,$xb0,$xb0		# "b3"
2218___
2219	($xb0,$xb1,$xb2,$xb3,$xt2)=($xb1,$xt2,$xb3,$xb0,$xb2);
2220$code.=<<___;
2221	vshufi32x4	\$0x44,$xb0,$xa0,$xt3	# "de-interlace" further
2222	vshufi32x4	\$0xee,$xb0,$xa0,$xb0
2223	vshufi32x4	\$0x44,$xb1,$xa1,$xa0
2224	vshufi32x4	\$0xee,$xb1,$xa1,$xb1
2225	vshufi32x4	\$0x44,$xb2,$xa2,$xa1
2226	vshufi32x4	\$0xee,$xb2,$xa2,$xb2
2227	vshufi32x4	\$0x44,$xb3,$xa3,$xa2
2228	vshufi32x4	\$0xee,$xb3,$xa3,$xb3
2229___
2230	($xa0,$xa1,$xa2,$xa3,$xt3)=($xt3,$xa0,$xa1,$xa2,$xa3);
2231$code.=<<___;
2232	vpaddd		@key[8],$xc0,$xc0
2233	vpaddd		@key[9],$xc1,$xc1
2234	vpaddd		@key[10],$xc2,$xc2
2235	vpaddd		@key[11],$xc3,$xc3
2236
2237	vpunpckldq	$xc1,$xc0,$xt2
2238	vpunpckldq	$xc3,$xc2,$xt3
2239	vpunpckhdq	$xc1,$xc0,$xc0
2240	vpunpckhdq	$xc3,$xc2,$xc2
2241	vpunpcklqdq	$xt3,$xt2,$xc1		# "c0"
2242	vpunpckhqdq	$xt3,$xt2,$xt2		# "c1"
2243	vpunpcklqdq	$xc2,$xc0,$xc3		# "c2"
2244	vpunpckhqdq	$xc2,$xc0,$xc0		# "c3"
2245___
2246	($xc0,$xc1,$xc2,$xc3,$xt2)=($xc1,$xt2,$xc3,$xc0,$xc2);
2247$code.=<<___;
2248	vpaddd		@key[12],$xd0,$xd0
2249	vpaddd		@key[13],$xd1,$xd1
2250	vpaddd		@key[14],$xd2,$xd2
2251	vpaddd		@key[15],$xd3,$xd3
2252
2253	vpunpckldq	$xd1,$xd0,$xt2
2254	vpunpckldq	$xd3,$xd2,$xt3
2255	vpunpckhdq	$xd1,$xd0,$xd0
2256	vpunpckhdq	$xd3,$xd2,$xd2
2257	vpunpcklqdq	$xt3,$xt2,$xd1		# "d0"
2258	vpunpckhqdq	$xt3,$xt2,$xt2		# "d1"
2259	vpunpcklqdq	$xd2,$xd0,$xd3		# "d2"
2260	vpunpckhqdq	$xd2,$xd0,$xd0		# "d3"
2261___
2262	($xd0,$xd1,$xd2,$xd3,$xt2)=($xd1,$xt2,$xd3,$xd0,$xd2);
2263$code.=<<___;
2264	vshufi32x4	\$0x44,$xd0,$xc0,$xt3	# "de-interlace" further
2265	vshufi32x4	\$0xee,$xd0,$xc0,$xd0
2266	vshufi32x4	\$0x44,$xd1,$xc1,$xc0
2267	vshufi32x4	\$0xee,$xd1,$xc1,$xd1
2268	vshufi32x4	\$0x44,$xd2,$xc2,$xc1
2269	vshufi32x4	\$0xee,$xd2,$xc2,$xd2
2270	vshufi32x4	\$0x44,$xd3,$xc3,$xc2
2271	vshufi32x4	\$0xee,$xd3,$xc3,$xd3
2272___
2273	($xc0,$xc1,$xc2,$xc3,$xt3)=($xt3,$xc0,$xc1,$xc2,$xc3);
2274$code.=<<___;
2275	vshufi32x4	\$0x88,$xc0,$xa0,$xt0	# "de-interlace" further
2276	vshufi32x4	\$0xdd,$xc0,$xa0,$xa0
2277	 vshufi32x4	\$0x88,$xd0,$xb0,$xc0
2278	 vshufi32x4	\$0xdd,$xd0,$xb0,$xd0
2279	vshufi32x4	\$0x88,$xc1,$xa1,$xt1
2280	vshufi32x4	\$0xdd,$xc1,$xa1,$xa1
2281	 vshufi32x4	\$0x88,$xd1,$xb1,$xc1
2282	 vshufi32x4	\$0xdd,$xd1,$xb1,$xd1
2283	vshufi32x4	\$0x88,$xc2,$xa2,$xt2
2284	vshufi32x4	\$0xdd,$xc2,$xa2,$xa2
2285	 vshufi32x4	\$0x88,$xd2,$xb2,$xc2
2286	 vshufi32x4	\$0xdd,$xd2,$xb2,$xd2
2287	vshufi32x4	\$0x88,$xc3,$xa3,$xt3
2288	vshufi32x4	\$0xdd,$xc3,$xa3,$xa3
2289	 vshufi32x4	\$0x88,$xd3,$xb3,$xc3
2290	 vshufi32x4	\$0xdd,$xd3,$xb3,$xd3
2291___
2292	($xa0,$xa1,$xa2,$xa3,$xb0,$xb1,$xb2,$xb3)=
2293	($xt0,$xt1,$xt2,$xt3,$xa0,$xa1,$xa2,$xa3);
2294
2295	($xa0,$xb0,$xc0,$xd0, $xa1,$xb1,$xc1,$xd1,
2296	 $xa2,$xb2,$xc2,$xd2, $xa3,$xb3,$xc3,$xd3) =
2297	($xa0,$xa1,$xa2,$xa3, $xb0,$xb1,$xb2,$xb3,
2298	 $xc0,$xc1,$xc2,$xc3, $xd0,$xd1,$xd2,$xd3);
2299$code.=<<___;
2300	cmp		\$64*16,$len
2301	jb		.Ltail16x
2302
2303	vpxord		0x00($inp),$xa0,$xa0	# xor with input
2304	vpxord		0x40($inp),$xb0,$xb0
2305	vpxord		0x80($inp),$xc0,$xc0
2306	vpxord		0xc0($inp),$xd0,$xd0
2307	vmovdqu32	$xa0,0x00($out)
2308	vmovdqu32	$xb0,0x40($out)
2309	vmovdqu32	$xc0,0x80($out)
2310	vmovdqu32	$xd0,0xc0($out)
2311
2312	vpxord		0x100($inp),$xa1,$xa1
2313	vpxord		0x140($inp),$xb1,$xb1
2314	vpxord		0x180($inp),$xc1,$xc1
2315	vpxord		0x1c0($inp),$xd1,$xd1
2316	vmovdqu32	$xa1,0x100($out)
2317	vmovdqu32	$xb1,0x140($out)
2318	vmovdqu32	$xc1,0x180($out)
2319	vmovdqu32	$xd1,0x1c0($out)
2320
2321	vpxord		0x200($inp),$xa2,$xa2
2322	vpxord		0x240($inp),$xb2,$xb2
2323	vpxord		0x280($inp),$xc2,$xc2
2324	vpxord		0x2c0($inp),$xd2,$xd2
2325	vmovdqu32	$xa2,0x200($out)
2326	vmovdqu32	$xb2,0x240($out)
2327	vmovdqu32	$xc2,0x280($out)
2328	vmovdqu32	$xd2,0x2c0($out)
2329
2330	vpxord		0x300($inp),$xa3,$xa3
2331	vpxord		0x340($inp),$xb3,$xb3
2332	vpxord		0x380($inp),$xc3,$xc3
2333	vpxord		0x3c0($inp),$xd3,$xd3
2334	lea		0x400($inp),$inp
2335	vmovdqu32	$xa3,0x300($out)
2336	vmovdqu32	$xb3,0x340($out)
2337	vmovdqu32	$xc3,0x380($out)
2338	vmovdqu32	$xd3,0x3c0($out)
2339	lea		0x400($out),$out
2340
2341	sub		\$64*16,$len
2342	jnz		.Loop_outer16x
2343
2344	jmp		.Ldone16x
2345
2346.align	32
2347.Ltail16x:
2348	xor		%r10,%r10
2349	sub		$inp,$out
2350	cmp		\$64*1,$len
2351	jb		.Less_than_64_16x
2352	vpxord		($inp),$xa0,$xa0	# xor with input
2353	vmovdqu32	$xa0,($out,$inp)
2354	je		.Ldone16x
2355	vmovdqa32	$xb0,$xa0
2356	lea		64($inp),$inp
2357
2358	cmp		\$64*2,$len
2359	jb		.Less_than_64_16x
2360	vpxord		($inp),$xb0,$xb0
2361	vmovdqu32	$xb0,($out,$inp)
2362	je		.Ldone16x
2363	vmovdqa32	$xc0,$xa0
2364	lea		64($inp),$inp
2365
2366	cmp		\$64*3,$len
2367	jb		.Less_than_64_16x
2368	vpxord		($inp),$xc0,$xc0
2369	vmovdqu32	$xc0,($out,$inp)
2370	je		.Ldone16x
2371	vmovdqa32	$xd0,$xa0
2372	lea		64($inp),$inp
2373
2374	cmp		\$64*4,$len
2375	jb		.Less_than_64_16x
2376	vpxord		($inp),$xd0,$xd0
2377	vmovdqu32	$xd0,($out,$inp)
2378	je		.Ldone16x
2379	vmovdqa32	$xa1,$xa0
2380	lea		64($inp),$inp
2381
2382	cmp		\$64*5,$len
2383	jb		.Less_than_64_16x
2384	vpxord		($inp),$xa1,$xa1
2385	vmovdqu32	$xa1,($out,$inp)
2386	je		.Ldone16x
2387	vmovdqa32	$xb1,$xa0
2388	lea		64($inp),$inp
2389
2390	cmp		\$64*6,$len
2391	jb		.Less_than_64_16x
2392	vpxord		($inp),$xb1,$xb1
2393	vmovdqu32	$xb1,($out,$inp)
2394	je		.Ldone16x
2395	vmovdqa32	$xc1,$xa0
2396	lea		64($inp),$inp
2397
2398	cmp		\$64*7,$len
2399	jb		.Less_than_64_16x
2400	vpxord		($inp),$xc1,$xc1
2401	vmovdqu32	$xc1,($out,$inp)
2402	je		.Ldone16x
2403	vmovdqa32	$xd1,$xa0
2404	lea		64($inp),$inp
2405
2406	cmp		\$64*8,$len
2407	jb		.Less_than_64_16x
2408	vpxord		($inp),$xd1,$xd1
2409	vmovdqu32	$xd1,($out,$inp)
2410	je		.Ldone16x
2411	vmovdqa32	$xa2,$xa0
2412	lea		64($inp),$inp
2413
2414	cmp		\$64*9,$len
2415	jb		.Less_than_64_16x
2416	vpxord		($inp),$xa2,$xa2
2417	vmovdqu32	$xa2,($out,$inp)
2418	je		.Ldone16x
2419	vmovdqa32	$xb2,$xa0
2420	lea		64($inp),$inp
2421
2422	cmp		\$64*10,$len
2423	jb		.Less_than_64_16x
2424	vpxord		($inp),$xb2,$xb2
2425	vmovdqu32	$xb2,($out,$inp)
2426	je		.Ldone16x
2427	vmovdqa32	$xc2,$xa0
2428	lea		64($inp),$inp
2429
2430	cmp		\$64*11,$len
2431	jb		.Less_than_64_16x
2432	vpxord		($inp),$xc2,$xc2
2433	vmovdqu32	$xc2,($out,$inp)
2434	je		.Ldone16x
2435	vmovdqa32	$xd2,$xa0
2436	lea		64($inp),$inp
2437
2438	cmp		\$64*12,$len
2439	jb		.Less_than_64_16x
2440	vpxord		($inp),$xd2,$xd2
2441	vmovdqu32	$xd2,($out,$inp)
2442	je		.Ldone16x
2443	vmovdqa32	$xa3,$xa0
2444	lea		64($inp),$inp
2445
2446	cmp		\$64*13,$len
2447	jb		.Less_than_64_16x
2448	vpxord		($inp),$xa3,$xa3
2449	vmovdqu32	$xa3,($out,$inp)
2450	je		.Ldone16x
2451	vmovdqa32	$xb3,$xa0
2452	lea		64($inp),$inp
2453
2454	cmp		\$64*14,$len
2455	jb		.Less_than_64_16x
2456	vpxord		($inp),$xb3,$xb3
2457	vmovdqu32	$xb3,($out,$inp)
2458	je		.Ldone16x
2459	vmovdqa32	$xc3,$xa0
2460	lea		64($inp),$inp
2461
2462	cmp		\$64*15,$len
2463	jb		.Less_than_64_16x
2464	vpxord		($inp),$xc3,$xc3
2465	vmovdqu32	$xc3,($out,$inp)
2466	je		.Ldone16x
2467	vmovdqa32	$xd3,$xa0
2468	lea		64($inp),$inp
2469
2470.Less_than_64_16x:
2471	vmovdqa32	$xa0,0x00(%rsp)
2472	lea		($out,$inp),$out
2473	and		\$63,$len
2474
2475.Loop_tail16x:
2476	movzb		($inp,%r10),%eax
2477	movzb		(%rsp,%r10),%ecx
2478	lea		1(%r10),%r10
2479	xor		%ecx,%eax
2480	mov		%al,-1($out,%r10)
2481	dec		$len
2482	jnz		.Loop_tail16x
2483
2484	vpxord		$xa0,$xa0,$xa0
2485	vmovdqa32	$xa0,0(%rsp)
2486
2487.Ldone16x:
2488	vzeroall
2489___
2490$code.=<<___	if ($win64);
2491	movaps		-0xa8(%r9),%xmm6
2492	movaps		-0x98(%r9),%xmm7
2493	movaps		-0x88(%r9),%xmm8
2494	movaps		-0x78(%r9),%xmm9
2495	movaps		-0x68(%r9),%xmm10
2496	movaps		-0x58(%r9),%xmm11
2497	movaps		-0x48(%r9),%xmm12
2498	movaps		-0x38(%r9),%xmm13
2499	movaps		-0x28(%r9),%xmm14
2500	movaps		-0x18(%r9),%xmm15
2501___
2502$code.=<<___;
2503	lea		(%r9),%rsp
2504.cfi_def_cfa_register	rsp
2505.L16x_epilogue:
2506	ret
2507.cfi_endproc
2508.size	ChaCha20_16x,.-ChaCha20_16x
2509___
2510}
2511
2512# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
2513#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
2514if ($win64) {
2515$rec="%rcx";
2516$frame="%rdx";
2517$context="%r8";
2518$disp="%r9";
2519
2520$code.=<<___;
2521.extern	__imp_RtlVirtualUnwind
2522.type	se_handler,\@abi-omnipotent
2523.align	16
2524se_handler:
2525	push	%rsi
2526	push	%rdi
2527	push	%rbx
2528	push	%rbp
2529	push	%r12
2530	push	%r13
2531	push	%r14
2532	push	%r15
2533	pushfq
2534	sub	\$64,%rsp
2535
2536	mov	120($context),%rax	# pull context->Rax
2537	mov	248($context),%rbx	# pull context->Rip
2538
2539	mov	8($disp),%rsi		# disp->ImageBase
2540	mov	56($disp),%r11		# disp->HandlerData
2541
2542	lea	.Lctr32_body(%rip),%r10
2543	cmp	%r10,%rbx		# context->Rip<.Lprologue
2544	jb	.Lcommon_seh_tail
2545
2546	mov	152($context),%rax	# pull context->Rsp
2547
2548	lea	.Lno_data(%rip),%r10	# epilogue label
2549	cmp	%r10,%rbx		# context->Rip>=.Lepilogue
2550	jae	.Lcommon_seh_tail
2551
2552	lea	64+24+48(%rax),%rax
2553
2554	mov	-8(%rax),%rbx
2555	mov	-16(%rax),%rbp
2556	mov	-24(%rax),%r12
2557	mov	-32(%rax),%r13
2558	mov	-40(%rax),%r14
2559	mov	-48(%rax),%r15
2560	mov	%rbx,144($context)	# restore context->Rbx
2561	mov	%rbp,160($context)	# restore context->Rbp
2562	mov	%r12,216($context)	# restore context->R12
2563	mov	%r13,224($context)	# restore context->R13
2564	mov	%r14,232($context)	# restore context->R14
2565	mov	%r15,240($context)	# restore context->R14
2566
2567.Lcommon_seh_tail:
2568	mov	8(%rax),%rdi
2569	mov	16(%rax),%rsi
2570	mov	%rax,152($context)	# restore context->Rsp
2571	mov	%rsi,168($context)	# restore context->Rsi
2572	mov	%rdi,176($context)	# restore context->Rdi
2573
2574	mov	40($disp),%rdi		# disp->ContextRecord
2575	mov	$context,%rsi		# context
2576	mov	\$154,%ecx		# sizeof(CONTEXT)
2577	.long	0xa548f3fc		# cld; rep movsq
2578
2579	mov	$disp,%rsi
2580	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
2581	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
2582	mov	0(%rsi),%r8		# arg3, disp->ControlPc
2583	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
2584	mov	40(%rsi),%r10		# disp->ContextRecord
2585	lea	56(%rsi),%r11		# &disp->HandlerData
2586	lea	24(%rsi),%r12		# &disp->EstablisherFrame
2587	mov	%r10,32(%rsp)		# arg5
2588	mov	%r11,40(%rsp)		# arg6
2589	mov	%r12,48(%rsp)		# arg7
2590	mov	%rcx,56(%rsp)		# arg8, (NULL)
2591	call	*__imp_RtlVirtualUnwind(%rip)
2592
2593	mov	\$1,%eax		# ExceptionContinueSearch
2594	add	\$64,%rsp
2595	popfq
2596	pop	%r15
2597	pop	%r14
2598	pop	%r13
2599	pop	%r12
2600	pop	%rbp
2601	pop	%rbx
2602	pop	%rdi
2603	pop	%rsi
2604	ret
2605.size	se_handler,.-se_handler
2606
2607.type	ssse3_handler,\@abi-omnipotent
2608.align	16
2609ssse3_handler:
2610	push	%rsi
2611	push	%rdi
2612	push	%rbx
2613	push	%rbp
2614	push	%r12
2615	push	%r13
2616	push	%r14
2617	push	%r15
2618	pushfq
2619	sub	\$64,%rsp
2620
2621	mov	120($context),%rax	# pull context->Rax
2622	mov	248($context),%rbx	# pull context->Rip
2623
2624	mov	8($disp),%rsi		# disp->ImageBase
2625	mov	56($disp),%r11		# disp->HandlerData
2626
2627	mov	0(%r11),%r10d		# HandlerData[0]
2628	lea	(%rsi,%r10),%r10	# prologue label
2629	cmp	%r10,%rbx		# context->Rip<prologue label
2630	jb	.Lcommon_seh_tail
2631
2632	mov	192($context),%rax	# pull context->R9
2633
2634	mov	4(%r11),%r10d		# HandlerData[1]
2635	lea	(%rsi,%r10),%r10	# epilogue label
2636	cmp	%r10,%rbx		# context->Rip>=epilogue label
2637	jae	.Lcommon_seh_tail
2638
2639	lea	-0x28(%rax),%rsi
2640	lea	512($context),%rdi	# &context.Xmm6
2641	mov	\$4,%ecx
2642	.long	0xa548f3fc		# cld; rep movsq
2643
2644	jmp	.Lcommon_seh_tail
2645.size	ssse3_handler,.-ssse3_handler
2646
2647.type	full_handler,\@abi-omnipotent
2648.align	16
2649full_handler:
2650	push	%rsi
2651	push	%rdi
2652	push	%rbx
2653	push	%rbp
2654	push	%r12
2655	push	%r13
2656	push	%r14
2657	push	%r15
2658	pushfq
2659	sub	\$64,%rsp
2660
2661	mov	120($context),%rax	# pull context->Rax
2662	mov	248($context),%rbx	# pull context->Rip
2663
2664	mov	8($disp),%rsi		# disp->ImageBase
2665	mov	56($disp),%r11		# disp->HandlerData
2666
2667	mov	0(%r11),%r10d		# HandlerData[0]
2668	lea	(%rsi,%r10),%r10	# prologue label
2669	cmp	%r10,%rbx		# context->Rip<prologue label
2670	jb	.Lcommon_seh_tail
2671
2672	mov	192($context),%rax	# pull context->R9
2673
2674	mov	4(%r11),%r10d		# HandlerData[1]
2675	lea	(%rsi,%r10),%r10	# epilogue label
2676	cmp	%r10,%rbx		# context->Rip>=epilogue label
2677	jae	.Lcommon_seh_tail
2678
2679	lea	-0xa8(%rax),%rsi
2680	lea	512($context),%rdi	# &context.Xmm6
2681	mov	\$20,%ecx
2682	.long	0xa548f3fc		# cld; rep movsq
2683
2684	jmp	.Lcommon_seh_tail
2685.size	full_handler,.-full_handler
2686
2687.section	.pdata
2688.align	4
2689	.rva	.LSEH_begin_ChaCha20_ctr32_nohw
2690	.rva	.LSEH_end_ChaCha20_ctr32_nohw
2691	.rva	.LSEH_info_ChaCha20_ctr32_nohw
2692
2693	.rva	.LSEH_begin_ChaCha20_ctr32_ssse3
2694	.rva	.LSEH_end_ChaCha20_ctr32_ssse3
2695	.rva	.LSEH_info_ChaCha20_ctr32_ssse3
2696
2697	.rva	.LSEH_begin_ChaCha20_ctr32_ssse3_4x
2698	.rva	.LSEH_end_ChaCha20_ctr32_ssse3_4x
2699	.rva	.LSEH_info_ChaCha20_ctr32_ssse3_4x
2700___
2701$code.=<<___ if ($avx>1);
2702	.rva	.LSEH_begin_ChaCha20_ctr32_avx2
2703	.rva	.LSEH_end_ChaCha20_ctr32_avx2
2704	.rva	.LSEH_info_ChaCha20_ctr32_avx2
2705___
2706$code.=<<___ if ($avx>2);
2707	.rva	.LSEH_begin_ChaCha20_avx512
2708	.rva	.LSEH_end_ChaCha20_avx512
2709	.rva	.LSEH_info_ChaCha20_avx512
2710
2711	.rva	.LSEH_begin_ChaCha20_16x
2712	.rva	.LSEH_end_ChaCha20_16x
2713	.rva	.LSEH_info_ChaCha20_16x
2714___
2715$code.=<<___;
2716.section	.xdata
2717.align	8
2718.LSEH_info_ChaCha20_ctr32_nohw:
2719	.byte	9,0,0,0
2720	.rva	se_handler
2721
2722.LSEH_info_ChaCha20_ctr32_ssse3:
2723	.byte	9,0,0,0
2724	.rva	ssse3_handler
2725	.rva	.Lssse3_body,.Lssse3_epilogue
2726
2727.LSEH_info_ChaCha20_ctr32_ssse3_4x:
2728	.byte	9,0,0,0
2729	.rva	full_handler
2730	.rva	.L4x_body,.L4x_epilogue
2731___
2732$code.=<<___ if ($avx>1);
2733.LSEH_info_ChaCha20_ctr32_avx2:
2734	.byte	9,0,0,0
2735	.rva	full_handler
2736	.rva	.L8x_body,.L8x_epilogue			# HandlerData[]
2737___
2738$code.=<<___ if ($avx>2);
2739.LSEH_info_ChaCha20_avx512:
2740	.byte	9,0,0,0
2741	.rva	ssse3_handler
2742	.rva	.Lavx512_body,.Lavx512_epilogue		# HandlerData[]
2743
2744.LSEH_info_ChaCha20_16x:
2745	.byte	9,0,0,0
2746	.rva	full_handler
2747	.rva	.L16x_body,.L16x_epilogue		# HandlerData[]
2748___
2749}
2750
2751foreach (split("\n",$code)) {
2752	s/\`([^\`]*)\`/eval $1/ge;
2753
2754	s/%x#%[yz]/%x/g;	# "down-shift"
2755
2756	print $_,"\n";
2757}
2758
2759close STDOUT or die "error closing STDOUT: $!";
2760