xref: /openssl/crypto/whrlpool/asm/wp-x86_64.pl (revision 7ed6de99)
1#! /usr/bin/env perl
2# Copyright 2005-2024 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the Apache License 2.0 (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9#
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. Rights for redistribution and usage in source and binary
13# forms are granted according to the License.
14# ====================================================================
15#
16# whirlpool_block for x86_64.
17#
18# 2500 cycles per 64-byte input block on AMD64, which is *identical*
19# to 32-bit MMX version executed on same CPU. So why did I bother?
20# Well, it's faster than gcc 3.3.2 generated code by over 50%, and
21# over 80% faster than PathScale 1.4, an "ambitious" commercial
22# compiler. Furthermore it surpasses gcc 3.4.3 by 170% and Sun Studio
23# 10 - by 360%[!]... What is it with x86_64 compilers? It's not the
24# first example when they fail to generate more optimal code, when
25# I believe they had *all* chances to...
26#
27# Note that register and stack frame layout are virtually identical
28# to 32-bit MMX version, except that %r8-15 are used instead of
29# %mm0-8. You can even notice that K[i] and S[i] are loaded to
30# %eax:%ebx as pair of 32-bit values and not as single 64-bit one.
31# This is done in order to avoid 64-bit shift penalties on Intel
32# EM64T core. Speaking of which! I bet it's possible to improve
33# Opteron performance by compressing the table to 2KB and replacing
34# unaligned references with complementary rotations [which would
35# incidentally replace lea instructions], but it would definitely
36# just "kill" EM64T, because it has only 1 shifter/rotator [against
37# 3 on Opteron] and which is *unacceptably* slow with 64-bit
38# operand.
39
40# $output is the last argument if it looks like a file (it has an extension)
41# $flavour is the first argument if it doesn't look like a file
42$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
43$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
44
45$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
46
47$0 =~ m/(.*[\/\\])[^\/\\]+$/; my $dir=$1; my $xlate;
48( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
49( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
50die "can't locate x86_64-xlate.pl";
51
52open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
53    or die "can't call $xlate: $!";
54*STDOUT=*OUT;
55
56sub L() { $code.=".byte	".join(',',@_)."\n"; }
57sub LL(){ $code.=".byte	".join(',',@_).",".join(',',@_)."\n"; }
58
59@mm=("%r8","%r9","%r10","%r11","%r12","%r13","%r14","%r15");
60
61$func="whirlpool_block";
62$table=".Ltable";
63
64$code=<<___;
65.text
66
67.globl	$func
68.type	$func,\@function,3
69.align	16
70$func:
71.cfi_startproc
72	mov	%rsp,%rax
73.cfi_def_cfa_register	%rax
74	push	%rbx
75.cfi_push	%rbx
76	push	%rbp
77.cfi_push	%rbp
78	push	%r12
79.cfi_push	%r12
80	push	%r13
81.cfi_push	%r13
82	push	%r14
83.cfi_push	%r14
84	push	%r15
85.cfi_push	%r15
86
87	sub	\$128+40,%rsp
88	and	\$-64,%rsp
89
90	lea	128(%rsp),%r10
91	mov	%rdi,0(%r10)		# save parameter block
92	mov	%rsi,8(%r10)
93	mov	%rdx,16(%r10)
94	mov	%rax,32(%r10)		# saved stack pointer
95.cfi_cfa_expression	%rsp+`128+32`,deref,+8
96.Lprologue:
97
98	mov	%r10,%rbx
99	lea	$table(%rip),%rbp
100
101	xor	%rcx,%rcx
102	xor	%rdx,%rdx
103___
104for($i=0;$i<8;$i++) { $code.="mov $i*8(%rdi),@mm[$i]\n"; }	# L=H
105$code.=".Louterloop:\n";
106for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; }	# K=L
107for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; }	# L^=inp
108for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; }	# S=L
109$code.=<<___;
110	xor	%rsi,%rsi
111	mov	%rsi,24(%rbx)		# zero round counter
112	jmp	.Lround
113.align	16
114.Lround:
115	mov	4096(%rbp,%rsi,8),@mm[0]	# rc[r]
116	mov	0(%rsp),%eax
117	mov	4(%rsp),%ebx
118	movz	%al,%ecx
119	movz	%ah,%edx
120___
121for($i=0;$i<8;$i++) {
122    my $func = ($i==0)? "mov" : "xor";
123    $code.=<<___;
124	shr	\$16,%eax
125	lea	(%rcx,%rcx),%rsi
126	movz	%al,%ecx
127	lea	(%rdx,%rdx),%rdi
128	movz	%ah,%edx
129	xor	0(%rbp,%rsi,8),@mm[0]
130	$func	7(%rbp,%rdi,8),@mm[1]
131	mov	$i*8+8(%rsp),%eax		# ($i+1)*8
132	lea	(%rcx,%rcx),%rsi
133	movz	%bl,%ecx
134	lea	(%rdx,%rdx),%rdi
135	movz	%bh,%edx
136	$func	6(%rbp,%rsi,8),@mm[2]
137	$func	5(%rbp,%rdi,8),@mm[3]
138	shr	\$16,%ebx
139	lea	(%rcx,%rcx),%rsi
140	movz	%bl,%ecx
141	lea	(%rdx,%rdx),%rdi
142	movz	%bh,%edx
143	$func	4(%rbp,%rsi,8),@mm[4]
144	$func	3(%rbp,%rdi,8),@mm[5]
145	mov	$i*8+8+4(%rsp),%ebx		# ($i+1)*8+4
146	lea	(%rcx,%rcx),%rsi
147	movz	%al,%ecx
148	lea	(%rdx,%rdx),%rdi
149	movz	%ah,%edx
150	$func	2(%rbp,%rsi,8),@mm[6]
151	$func	1(%rbp,%rdi,8),@mm[7]
152___
153    push(@mm,shift(@mm));
154}
155for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rsp)\n"; }	# K=L
156for($i=0;$i<8;$i++) {
157    $code.=<<___;
158	shr	\$16,%eax
159	lea	(%rcx,%rcx),%rsi
160	movz	%al,%ecx
161	lea	(%rdx,%rdx),%rdi
162	movz	%ah,%edx
163	xor	0(%rbp,%rsi,8),@mm[0]
164	xor	7(%rbp,%rdi,8),@mm[1]
165	`"mov	64+$i*8+8(%rsp),%eax"	if($i<7);`	# 64+($i+1)*8
166	lea	(%rcx,%rcx),%rsi
167	movz	%bl,%ecx
168	lea	(%rdx,%rdx),%rdi
169	movz	%bh,%edx
170	xor	6(%rbp,%rsi,8),@mm[2]
171	xor	5(%rbp,%rdi,8),@mm[3]
172	shr	\$16,%ebx
173	lea	(%rcx,%rcx),%rsi
174	movz	%bl,%ecx
175	lea	(%rdx,%rdx),%rdi
176	movz	%bh,%edx
177	xor	4(%rbp,%rsi,8),@mm[4]
178	xor	3(%rbp,%rdi,8),@mm[5]
179	`"mov	64+$i*8+8+4(%rsp),%ebx"	if($i<7);`	# 64+($i+1)*8+4
180	lea	(%rcx,%rcx),%rsi
181	movz	%al,%ecx
182	lea	(%rdx,%rdx),%rdi
183	movz	%ah,%edx
184	xor	2(%rbp,%rsi,8),@mm[6]
185	xor	1(%rbp,%rdi,8),@mm[7]
186___
187    push(@mm,shift(@mm));
188}
189$code.=<<___;
190	lea	128(%rsp),%rbx
191	mov	24(%rbx),%rsi		# pull round counter
192	add	\$1,%rsi
193	cmp	\$10,%rsi
194	je	.Lroundsdone
195
196	mov	%rsi,24(%rbx)		# update round counter
197___
198for($i=0;$i<8;$i++) { $code.="mov @mm[$i],64+$i*8(%rsp)\n"; }	# S=L
199$code.=<<___;
200	jmp	.Lround
201.align	16
202.Lroundsdone:
203	mov	0(%rbx),%rdi		# reload argument block
204	mov	8(%rbx),%rsi
205	mov	16(%rbx),%rax
206___
207for($i=0;$i<8;$i++) { $code.="xor $i*8(%rsi),@mm[$i]\n"; }	# L^=inp
208for($i=0;$i<8;$i++) { $code.="xor $i*8(%rdi),@mm[$i]\n"; }	# L^=H
209for($i=0;$i<8;$i++) { $code.="mov @mm[$i],$i*8(%rdi)\n"; }	# H=L
210$code.=<<___;
211	lea	64(%rsi),%rsi		# inp+=64
212	sub	\$1,%rax		# num--
213	jz	.Lalldone
214	mov	%rsi,8(%rbx)		# update parameter block
215	mov	%rax,16(%rbx)
216	jmp	.Louterloop
217.Lalldone:
218	mov	32(%rbx),%rsi		# restore saved pointer
219.cfi_def_cfa	%rsi,8
220	mov	-48(%rsi),%r15
221.cfi_restore	%r15
222	mov	-40(%rsi),%r14
223.cfi_restore	%r14
224	mov	-32(%rsi),%r13
225.cfi_restore	%r13
226	mov	-24(%rsi),%r12
227.cfi_restore	%r12
228	mov	-16(%rsi),%rbp
229.cfi_restore	%rbp
230	mov	-8(%rsi),%rbx
231.cfi_restore	%rbx
232	lea	(%rsi),%rsp
233.cfi_def_cfa_register	%rsp
234.Lepilogue:
235	ret
236.cfi_endproc
237.size	$func,.-$func
238
239.section .rodata align=64
240.align	64
241.type	$table,\@object
242$table:
243___
244	&LL(0x18,0x18,0x60,0x18,0xc0,0x78,0x30,0xd8);
245	&LL(0x23,0x23,0x8c,0x23,0x05,0xaf,0x46,0x26);
246	&LL(0xc6,0xc6,0x3f,0xc6,0x7e,0xf9,0x91,0xb8);
247	&LL(0xe8,0xe8,0x87,0xe8,0x13,0x6f,0xcd,0xfb);
248	&LL(0x87,0x87,0x26,0x87,0x4c,0xa1,0x13,0xcb);
249	&LL(0xb8,0xb8,0xda,0xb8,0xa9,0x62,0x6d,0x11);
250	&LL(0x01,0x01,0x04,0x01,0x08,0x05,0x02,0x09);
251	&LL(0x4f,0x4f,0x21,0x4f,0x42,0x6e,0x9e,0x0d);
252	&LL(0x36,0x36,0xd8,0x36,0xad,0xee,0x6c,0x9b);
253	&LL(0xa6,0xa6,0xa2,0xa6,0x59,0x04,0x51,0xff);
254	&LL(0xd2,0xd2,0x6f,0xd2,0xde,0xbd,0xb9,0x0c);
255	&LL(0xf5,0xf5,0xf3,0xf5,0xfb,0x06,0xf7,0x0e);
256	&LL(0x79,0x79,0xf9,0x79,0xef,0x80,0xf2,0x96);
257	&LL(0x6f,0x6f,0xa1,0x6f,0x5f,0xce,0xde,0x30);
258	&LL(0x91,0x91,0x7e,0x91,0xfc,0xef,0x3f,0x6d);
259	&LL(0x52,0x52,0x55,0x52,0xaa,0x07,0xa4,0xf8);
260	&LL(0x60,0x60,0x9d,0x60,0x27,0xfd,0xc0,0x47);
261	&LL(0xbc,0xbc,0xca,0xbc,0x89,0x76,0x65,0x35);
262	&LL(0x9b,0x9b,0x56,0x9b,0xac,0xcd,0x2b,0x37);
263	&LL(0x8e,0x8e,0x02,0x8e,0x04,0x8c,0x01,0x8a);
264	&LL(0xa3,0xa3,0xb6,0xa3,0x71,0x15,0x5b,0xd2);
265	&LL(0x0c,0x0c,0x30,0x0c,0x60,0x3c,0x18,0x6c);
266	&LL(0x7b,0x7b,0xf1,0x7b,0xff,0x8a,0xf6,0x84);
267	&LL(0x35,0x35,0xd4,0x35,0xb5,0xe1,0x6a,0x80);
268	&LL(0x1d,0x1d,0x74,0x1d,0xe8,0x69,0x3a,0xf5);
269	&LL(0xe0,0xe0,0xa7,0xe0,0x53,0x47,0xdd,0xb3);
270	&LL(0xd7,0xd7,0x7b,0xd7,0xf6,0xac,0xb3,0x21);
271	&LL(0xc2,0xc2,0x2f,0xc2,0x5e,0xed,0x99,0x9c);
272	&LL(0x2e,0x2e,0xb8,0x2e,0x6d,0x96,0x5c,0x43);
273	&LL(0x4b,0x4b,0x31,0x4b,0x62,0x7a,0x96,0x29);
274	&LL(0xfe,0xfe,0xdf,0xfe,0xa3,0x21,0xe1,0x5d);
275	&LL(0x57,0x57,0x41,0x57,0x82,0x16,0xae,0xd5);
276	&LL(0x15,0x15,0x54,0x15,0xa8,0x41,0x2a,0xbd);
277	&LL(0x77,0x77,0xc1,0x77,0x9f,0xb6,0xee,0xe8);
278	&LL(0x37,0x37,0xdc,0x37,0xa5,0xeb,0x6e,0x92);
279	&LL(0xe5,0xe5,0xb3,0xe5,0x7b,0x56,0xd7,0x9e);
280	&LL(0x9f,0x9f,0x46,0x9f,0x8c,0xd9,0x23,0x13);
281	&LL(0xf0,0xf0,0xe7,0xf0,0xd3,0x17,0xfd,0x23);
282	&LL(0x4a,0x4a,0x35,0x4a,0x6a,0x7f,0x94,0x20);
283	&LL(0xda,0xda,0x4f,0xda,0x9e,0x95,0xa9,0x44);
284	&LL(0x58,0x58,0x7d,0x58,0xfa,0x25,0xb0,0xa2);
285	&LL(0xc9,0xc9,0x03,0xc9,0x06,0xca,0x8f,0xcf);
286	&LL(0x29,0x29,0xa4,0x29,0x55,0x8d,0x52,0x7c);
287	&LL(0x0a,0x0a,0x28,0x0a,0x50,0x22,0x14,0x5a);
288	&LL(0xb1,0xb1,0xfe,0xb1,0xe1,0x4f,0x7f,0x50);
289	&LL(0xa0,0xa0,0xba,0xa0,0x69,0x1a,0x5d,0xc9);
290	&LL(0x6b,0x6b,0xb1,0x6b,0x7f,0xda,0xd6,0x14);
291	&LL(0x85,0x85,0x2e,0x85,0x5c,0xab,0x17,0xd9);
292	&LL(0xbd,0xbd,0xce,0xbd,0x81,0x73,0x67,0x3c);
293	&LL(0x5d,0x5d,0x69,0x5d,0xd2,0x34,0xba,0x8f);
294	&LL(0x10,0x10,0x40,0x10,0x80,0x50,0x20,0x90);
295	&LL(0xf4,0xf4,0xf7,0xf4,0xf3,0x03,0xf5,0x07);
296	&LL(0xcb,0xcb,0x0b,0xcb,0x16,0xc0,0x8b,0xdd);
297	&LL(0x3e,0x3e,0xf8,0x3e,0xed,0xc6,0x7c,0xd3);
298	&LL(0x05,0x05,0x14,0x05,0x28,0x11,0x0a,0x2d);
299	&LL(0x67,0x67,0x81,0x67,0x1f,0xe6,0xce,0x78);
300	&LL(0xe4,0xe4,0xb7,0xe4,0x73,0x53,0xd5,0x97);
301	&LL(0x27,0x27,0x9c,0x27,0x25,0xbb,0x4e,0x02);
302	&LL(0x41,0x41,0x19,0x41,0x32,0x58,0x82,0x73);
303	&LL(0x8b,0x8b,0x16,0x8b,0x2c,0x9d,0x0b,0xa7);
304	&LL(0xa7,0xa7,0xa6,0xa7,0x51,0x01,0x53,0xf6);
305	&LL(0x7d,0x7d,0xe9,0x7d,0xcf,0x94,0xfa,0xb2);
306	&LL(0x95,0x95,0x6e,0x95,0xdc,0xfb,0x37,0x49);
307	&LL(0xd8,0xd8,0x47,0xd8,0x8e,0x9f,0xad,0x56);
308	&LL(0xfb,0xfb,0xcb,0xfb,0x8b,0x30,0xeb,0x70);
309	&LL(0xee,0xee,0x9f,0xee,0x23,0x71,0xc1,0xcd);
310	&LL(0x7c,0x7c,0xed,0x7c,0xc7,0x91,0xf8,0xbb);
311	&LL(0x66,0x66,0x85,0x66,0x17,0xe3,0xcc,0x71);
312	&LL(0xdd,0xdd,0x53,0xdd,0xa6,0x8e,0xa7,0x7b);
313	&LL(0x17,0x17,0x5c,0x17,0xb8,0x4b,0x2e,0xaf);
314	&LL(0x47,0x47,0x01,0x47,0x02,0x46,0x8e,0x45);
315	&LL(0x9e,0x9e,0x42,0x9e,0x84,0xdc,0x21,0x1a);
316	&LL(0xca,0xca,0x0f,0xca,0x1e,0xc5,0x89,0xd4);
317	&LL(0x2d,0x2d,0xb4,0x2d,0x75,0x99,0x5a,0x58);
318	&LL(0xbf,0xbf,0xc6,0xbf,0x91,0x79,0x63,0x2e);
319	&LL(0x07,0x07,0x1c,0x07,0x38,0x1b,0x0e,0x3f);
320	&LL(0xad,0xad,0x8e,0xad,0x01,0x23,0x47,0xac);
321	&LL(0x5a,0x5a,0x75,0x5a,0xea,0x2f,0xb4,0xb0);
322	&LL(0x83,0x83,0x36,0x83,0x6c,0xb5,0x1b,0xef);
323	&LL(0x33,0x33,0xcc,0x33,0x85,0xff,0x66,0xb6);
324	&LL(0x63,0x63,0x91,0x63,0x3f,0xf2,0xc6,0x5c);
325	&LL(0x02,0x02,0x08,0x02,0x10,0x0a,0x04,0x12);
326	&LL(0xaa,0xaa,0x92,0xaa,0x39,0x38,0x49,0x93);
327	&LL(0x71,0x71,0xd9,0x71,0xaf,0xa8,0xe2,0xde);
328	&LL(0xc8,0xc8,0x07,0xc8,0x0e,0xcf,0x8d,0xc6);
329	&LL(0x19,0x19,0x64,0x19,0xc8,0x7d,0x32,0xd1);
330	&LL(0x49,0x49,0x39,0x49,0x72,0x70,0x92,0x3b);
331	&LL(0xd9,0xd9,0x43,0xd9,0x86,0x9a,0xaf,0x5f);
332	&LL(0xf2,0xf2,0xef,0xf2,0xc3,0x1d,0xf9,0x31);
333	&LL(0xe3,0xe3,0xab,0xe3,0x4b,0x48,0xdb,0xa8);
334	&LL(0x5b,0x5b,0x71,0x5b,0xe2,0x2a,0xb6,0xb9);
335	&LL(0x88,0x88,0x1a,0x88,0x34,0x92,0x0d,0xbc);
336	&LL(0x9a,0x9a,0x52,0x9a,0xa4,0xc8,0x29,0x3e);
337	&LL(0x26,0x26,0x98,0x26,0x2d,0xbe,0x4c,0x0b);
338	&LL(0x32,0x32,0xc8,0x32,0x8d,0xfa,0x64,0xbf);
339	&LL(0xb0,0xb0,0xfa,0xb0,0xe9,0x4a,0x7d,0x59);
340	&LL(0xe9,0xe9,0x83,0xe9,0x1b,0x6a,0xcf,0xf2);
341	&LL(0x0f,0x0f,0x3c,0x0f,0x78,0x33,0x1e,0x77);
342	&LL(0xd5,0xd5,0x73,0xd5,0xe6,0xa6,0xb7,0x33);
343	&LL(0x80,0x80,0x3a,0x80,0x74,0xba,0x1d,0xf4);
344	&LL(0xbe,0xbe,0xc2,0xbe,0x99,0x7c,0x61,0x27);
345	&LL(0xcd,0xcd,0x13,0xcd,0x26,0xde,0x87,0xeb);
346	&LL(0x34,0x34,0xd0,0x34,0xbd,0xe4,0x68,0x89);
347	&LL(0x48,0x48,0x3d,0x48,0x7a,0x75,0x90,0x32);
348	&LL(0xff,0xff,0xdb,0xff,0xab,0x24,0xe3,0x54);
349	&LL(0x7a,0x7a,0xf5,0x7a,0xf7,0x8f,0xf4,0x8d);
350	&LL(0x90,0x90,0x7a,0x90,0xf4,0xea,0x3d,0x64);
351	&LL(0x5f,0x5f,0x61,0x5f,0xc2,0x3e,0xbe,0x9d);
352	&LL(0x20,0x20,0x80,0x20,0x1d,0xa0,0x40,0x3d);
353	&LL(0x68,0x68,0xbd,0x68,0x67,0xd5,0xd0,0x0f);
354	&LL(0x1a,0x1a,0x68,0x1a,0xd0,0x72,0x34,0xca);
355	&LL(0xae,0xae,0x82,0xae,0x19,0x2c,0x41,0xb7);
356	&LL(0xb4,0xb4,0xea,0xb4,0xc9,0x5e,0x75,0x7d);
357	&LL(0x54,0x54,0x4d,0x54,0x9a,0x19,0xa8,0xce);
358	&LL(0x93,0x93,0x76,0x93,0xec,0xe5,0x3b,0x7f);
359	&LL(0x22,0x22,0x88,0x22,0x0d,0xaa,0x44,0x2f);
360	&LL(0x64,0x64,0x8d,0x64,0x07,0xe9,0xc8,0x63);
361	&LL(0xf1,0xf1,0xe3,0xf1,0xdb,0x12,0xff,0x2a);
362	&LL(0x73,0x73,0xd1,0x73,0xbf,0xa2,0xe6,0xcc);
363	&LL(0x12,0x12,0x48,0x12,0x90,0x5a,0x24,0x82);
364	&LL(0x40,0x40,0x1d,0x40,0x3a,0x5d,0x80,0x7a);
365	&LL(0x08,0x08,0x20,0x08,0x40,0x28,0x10,0x48);
366	&LL(0xc3,0xc3,0x2b,0xc3,0x56,0xe8,0x9b,0x95);
367	&LL(0xec,0xec,0x97,0xec,0x33,0x7b,0xc5,0xdf);
368	&LL(0xdb,0xdb,0x4b,0xdb,0x96,0x90,0xab,0x4d);
369	&LL(0xa1,0xa1,0xbe,0xa1,0x61,0x1f,0x5f,0xc0);
370	&LL(0x8d,0x8d,0x0e,0x8d,0x1c,0x83,0x07,0x91);
371	&LL(0x3d,0x3d,0xf4,0x3d,0xf5,0xc9,0x7a,0xc8);
372	&LL(0x97,0x97,0x66,0x97,0xcc,0xf1,0x33,0x5b);
373	&LL(0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
374	&LL(0xcf,0xcf,0x1b,0xcf,0x36,0xd4,0x83,0xf9);
375	&LL(0x2b,0x2b,0xac,0x2b,0x45,0x87,0x56,0x6e);
376	&LL(0x76,0x76,0xc5,0x76,0x97,0xb3,0xec,0xe1);
377	&LL(0x82,0x82,0x32,0x82,0x64,0xb0,0x19,0xe6);
378	&LL(0xd6,0xd6,0x7f,0xd6,0xfe,0xa9,0xb1,0x28);
379	&LL(0x1b,0x1b,0x6c,0x1b,0xd8,0x77,0x36,0xc3);
380	&LL(0xb5,0xb5,0xee,0xb5,0xc1,0x5b,0x77,0x74);
381	&LL(0xaf,0xaf,0x86,0xaf,0x11,0x29,0x43,0xbe);
382	&LL(0x6a,0x6a,0xb5,0x6a,0x77,0xdf,0xd4,0x1d);
383	&LL(0x50,0x50,0x5d,0x50,0xba,0x0d,0xa0,0xea);
384	&LL(0x45,0x45,0x09,0x45,0x12,0x4c,0x8a,0x57);
385	&LL(0xf3,0xf3,0xeb,0xf3,0xcb,0x18,0xfb,0x38);
386	&LL(0x30,0x30,0xc0,0x30,0x9d,0xf0,0x60,0xad);
387	&LL(0xef,0xef,0x9b,0xef,0x2b,0x74,0xc3,0xc4);
388	&LL(0x3f,0x3f,0xfc,0x3f,0xe5,0xc3,0x7e,0xda);
389	&LL(0x55,0x55,0x49,0x55,0x92,0x1c,0xaa,0xc7);
390	&LL(0xa2,0xa2,0xb2,0xa2,0x79,0x10,0x59,0xdb);
391	&LL(0xea,0xea,0x8f,0xea,0x03,0x65,0xc9,0xe9);
392	&LL(0x65,0x65,0x89,0x65,0x0f,0xec,0xca,0x6a);
393	&LL(0xba,0xba,0xd2,0xba,0xb9,0x68,0x69,0x03);
394	&LL(0x2f,0x2f,0xbc,0x2f,0x65,0x93,0x5e,0x4a);
395	&LL(0xc0,0xc0,0x27,0xc0,0x4e,0xe7,0x9d,0x8e);
396	&LL(0xde,0xde,0x5f,0xde,0xbe,0x81,0xa1,0x60);
397	&LL(0x1c,0x1c,0x70,0x1c,0xe0,0x6c,0x38,0xfc);
398	&LL(0xfd,0xfd,0xd3,0xfd,0xbb,0x2e,0xe7,0x46);
399	&LL(0x4d,0x4d,0x29,0x4d,0x52,0x64,0x9a,0x1f);
400	&LL(0x92,0x92,0x72,0x92,0xe4,0xe0,0x39,0x76);
401	&LL(0x75,0x75,0xc9,0x75,0x8f,0xbc,0xea,0xfa);
402	&LL(0x06,0x06,0x18,0x06,0x30,0x1e,0x0c,0x36);
403	&LL(0x8a,0x8a,0x12,0x8a,0x24,0x98,0x09,0xae);
404	&LL(0xb2,0xb2,0xf2,0xb2,0xf9,0x40,0x79,0x4b);
405	&LL(0xe6,0xe6,0xbf,0xe6,0x63,0x59,0xd1,0x85);
406	&LL(0x0e,0x0e,0x38,0x0e,0x70,0x36,0x1c,0x7e);
407	&LL(0x1f,0x1f,0x7c,0x1f,0xf8,0x63,0x3e,0xe7);
408	&LL(0x62,0x62,0x95,0x62,0x37,0xf7,0xc4,0x55);
409	&LL(0xd4,0xd4,0x77,0xd4,0xee,0xa3,0xb5,0x3a);
410	&LL(0xa8,0xa8,0x9a,0xa8,0x29,0x32,0x4d,0x81);
411	&LL(0x96,0x96,0x62,0x96,0xc4,0xf4,0x31,0x52);
412	&LL(0xf9,0xf9,0xc3,0xf9,0x9b,0x3a,0xef,0x62);
413	&LL(0xc5,0xc5,0x33,0xc5,0x66,0xf6,0x97,0xa3);
414	&LL(0x25,0x25,0x94,0x25,0x35,0xb1,0x4a,0x10);
415	&LL(0x59,0x59,0x79,0x59,0xf2,0x20,0xb2,0xab);
416	&LL(0x84,0x84,0x2a,0x84,0x54,0xae,0x15,0xd0);
417	&LL(0x72,0x72,0xd5,0x72,0xb7,0xa7,0xe4,0xc5);
418	&LL(0x39,0x39,0xe4,0x39,0xd5,0xdd,0x72,0xec);
419	&LL(0x4c,0x4c,0x2d,0x4c,0x5a,0x61,0x98,0x16);
420	&LL(0x5e,0x5e,0x65,0x5e,0xca,0x3b,0xbc,0x94);
421	&LL(0x78,0x78,0xfd,0x78,0xe7,0x85,0xf0,0x9f);
422	&LL(0x38,0x38,0xe0,0x38,0xdd,0xd8,0x70,0xe5);
423	&LL(0x8c,0x8c,0x0a,0x8c,0x14,0x86,0x05,0x98);
424	&LL(0xd1,0xd1,0x63,0xd1,0xc6,0xb2,0xbf,0x17);
425	&LL(0xa5,0xa5,0xae,0xa5,0x41,0x0b,0x57,0xe4);
426	&LL(0xe2,0xe2,0xaf,0xe2,0x43,0x4d,0xd9,0xa1);
427	&LL(0x61,0x61,0x99,0x61,0x2f,0xf8,0xc2,0x4e);
428	&LL(0xb3,0xb3,0xf6,0xb3,0xf1,0x45,0x7b,0x42);
429	&LL(0x21,0x21,0x84,0x21,0x15,0xa5,0x42,0x34);
430	&LL(0x9c,0x9c,0x4a,0x9c,0x94,0xd6,0x25,0x08);
431	&LL(0x1e,0x1e,0x78,0x1e,0xf0,0x66,0x3c,0xee);
432	&LL(0x43,0x43,0x11,0x43,0x22,0x52,0x86,0x61);
433	&LL(0xc7,0xc7,0x3b,0xc7,0x76,0xfc,0x93,0xb1);
434	&LL(0xfc,0xfc,0xd7,0xfc,0xb3,0x2b,0xe5,0x4f);
435	&LL(0x04,0x04,0x10,0x04,0x20,0x14,0x08,0x24);
436	&LL(0x51,0x51,0x59,0x51,0xb2,0x08,0xa2,0xe3);
437	&LL(0x99,0x99,0x5e,0x99,0xbc,0xc7,0x2f,0x25);
438	&LL(0x6d,0x6d,0xa9,0x6d,0x4f,0xc4,0xda,0x22);
439	&LL(0x0d,0x0d,0x34,0x0d,0x68,0x39,0x1a,0x65);
440	&LL(0xfa,0xfa,0xcf,0xfa,0x83,0x35,0xe9,0x79);
441	&LL(0xdf,0xdf,0x5b,0xdf,0xb6,0x84,0xa3,0x69);
442	&LL(0x7e,0x7e,0xe5,0x7e,0xd7,0x9b,0xfc,0xa9);
443	&LL(0x24,0x24,0x90,0x24,0x3d,0xb4,0x48,0x19);
444	&LL(0x3b,0x3b,0xec,0x3b,0xc5,0xd7,0x76,0xfe);
445	&LL(0xab,0xab,0x96,0xab,0x31,0x3d,0x4b,0x9a);
446	&LL(0xce,0xce,0x1f,0xce,0x3e,0xd1,0x81,0xf0);
447	&LL(0x11,0x11,0x44,0x11,0x88,0x55,0x22,0x99);
448	&LL(0x8f,0x8f,0x06,0x8f,0x0c,0x89,0x03,0x83);
449	&LL(0x4e,0x4e,0x25,0x4e,0x4a,0x6b,0x9c,0x04);
450	&LL(0xb7,0xb7,0xe6,0xb7,0xd1,0x51,0x73,0x66);
451	&LL(0xeb,0xeb,0x8b,0xeb,0x0b,0x60,0xcb,0xe0);
452	&LL(0x3c,0x3c,0xf0,0x3c,0xfd,0xcc,0x78,0xc1);
453	&LL(0x81,0x81,0x3e,0x81,0x7c,0xbf,0x1f,0xfd);
454	&LL(0x94,0x94,0x6a,0x94,0xd4,0xfe,0x35,0x40);
455	&LL(0xf7,0xf7,0xfb,0xf7,0xeb,0x0c,0xf3,0x1c);
456	&LL(0xb9,0xb9,0xde,0xb9,0xa1,0x67,0x6f,0x18);
457	&LL(0x13,0x13,0x4c,0x13,0x98,0x5f,0x26,0x8b);
458	&LL(0x2c,0x2c,0xb0,0x2c,0x7d,0x9c,0x58,0x51);
459	&LL(0xd3,0xd3,0x6b,0xd3,0xd6,0xb8,0xbb,0x05);
460	&LL(0xe7,0xe7,0xbb,0xe7,0x6b,0x5c,0xd3,0x8c);
461	&LL(0x6e,0x6e,0xa5,0x6e,0x57,0xcb,0xdc,0x39);
462	&LL(0xc4,0xc4,0x37,0xc4,0x6e,0xf3,0x95,0xaa);
463	&LL(0x03,0x03,0x0c,0x03,0x18,0x0f,0x06,0x1b);
464	&LL(0x56,0x56,0x45,0x56,0x8a,0x13,0xac,0xdc);
465	&LL(0x44,0x44,0x0d,0x44,0x1a,0x49,0x88,0x5e);
466	&LL(0x7f,0x7f,0xe1,0x7f,0xdf,0x9e,0xfe,0xa0);
467	&LL(0xa9,0xa9,0x9e,0xa9,0x21,0x37,0x4f,0x88);
468	&LL(0x2a,0x2a,0xa8,0x2a,0x4d,0x82,0x54,0x67);
469	&LL(0xbb,0xbb,0xd6,0xbb,0xb1,0x6d,0x6b,0x0a);
470	&LL(0xc1,0xc1,0x23,0xc1,0x46,0xe2,0x9f,0x87);
471	&LL(0x53,0x53,0x51,0x53,0xa2,0x02,0xa6,0xf1);
472	&LL(0xdc,0xdc,0x57,0xdc,0xae,0x8b,0xa5,0x72);
473	&LL(0x0b,0x0b,0x2c,0x0b,0x58,0x27,0x16,0x53);
474	&LL(0x9d,0x9d,0x4e,0x9d,0x9c,0xd3,0x27,0x01);
475	&LL(0x6c,0x6c,0xad,0x6c,0x47,0xc1,0xd8,0x2b);
476	&LL(0x31,0x31,0xc4,0x31,0x95,0xf5,0x62,0xa4);
477	&LL(0x74,0x74,0xcd,0x74,0x87,0xb9,0xe8,0xf3);
478	&LL(0xf6,0xf6,0xff,0xf6,0xe3,0x09,0xf1,0x15);
479	&LL(0x46,0x46,0x05,0x46,0x0a,0x43,0x8c,0x4c);
480	&LL(0xac,0xac,0x8a,0xac,0x09,0x26,0x45,0xa5);
481	&LL(0x89,0x89,0x1e,0x89,0x3c,0x97,0x0f,0xb5);
482	&LL(0x14,0x14,0x50,0x14,0xa0,0x44,0x28,0xb4);
483	&LL(0xe1,0xe1,0xa3,0xe1,0x5b,0x42,0xdf,0xba);
484	&LL(0x16,0x16,0x58,0x16,0xb0,0x4e,0x2c,0xa6);
485	&LL(0x3a,0x3a,0xe8,0x3a,0xcd,0xd2,0x74,0xf7);
486	&LL(0x69,0x69,0xb9,0x69,0x6f,0xd0,0xd2,0x06);
487	&LL(0x09,0x09,0x24,0x09,0x48,0x2d,0x12,0x41);
488	&LL(0x70,0x70,0xdd,0x70,0xa7,0xad,0xe0,0xd7);
489	&LL(0xb6,0xb6,0xe2,0xb6,0xd9,0x54,0x71,0x6f);
490	&LL(0xd0,0xd0,0x67,0xd0,0xce,0xb7,0xbd,0x1e);
491	&LL(0xed,0xed,0x93,0xed,0x3b,0x7e,0xc7,0xd6);
492	&LL(0xcc,0xcc,0x17,0xcc,0x2e,0xdb,0x85,0xe2);
493	&LL(0x42,0x42,0x15,0x42,0x2a,0x57,0x84,0x68);
494	&LL(0x98,0x98,0x5a,0x98,0xb4,0xc2,0x2d,0x2c);
495	&LL(0xa4,0xa4,0xaa,0xa4,0x49,0x0e,0x55,0xed);
496	&LL(0x28,0x28,0xa0,0x28,0x5d,0x88,0x50,0x75);
497	&LL(0x5c,0x5c,0x6d,0x5c,0xda,0x31,0xb8,0x86);
498	&LL(0xf8,0xf8,0xc7,0xf8,0x93,0x3f,0xed,0x6b);
499	&LL(0x86,0x86,0x22,0x86,0x44,0xa4,0x11,0xc2);
500
501	&L(0x18,0x23,0xc6,0xe8,0x87,0xb8,0x01,0x4f);	# rc[ROUNDS]
502	&L(0x36,0xa6,0xd2,0xf5,0x79,0x6f,0x91,0x52);
503	&L(0x60,0xbc,0x9b,0x8e,0xa3,0x0c,0x7b,0x35);
504	&L(0x1d,0xe0,0xd7,0xc2,0x2e,0x4b,0xfe,0x57);
505	&L(0x15,0x77,0x37,0xe5,0x9f,0xf0,0x4a,0xda);
506	&L(0x58,0xc9,0x29,0x0a,0xb1,0xa0,0x6b,0x85);
507	&L(0xbd,0x5d,0x10,0xf4,0xcb,0x3e,0x05,0x67);
508	&L(0xe4,0x27,0x41,0x8b,0xa7,0x7d,0x95,0xd8);
509	&L(0xfb,0xee,0x7c,0x66,0xdd,0x17,0x47,0x9e);
510	&L(0xca,0x2d,0xbf,0x07,0xad,0x5a,0x83,0x33);
511
512# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
513#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
514if ($win64) {
515$rec="%rcx";
516$frame="%rdx";
517$context="%r8";
518$disp="%r9";
519
520$code.=<<___;
521.extern	__imp_RtlVirtualUnwind
522.type	se_handler,\@abi-omnipotent
523.align	16
524se_handler:
525	push	%rsi
526	push	%rdi
527	push	%rbx
528	push	%rbp
529	push	%r12
530	push	%r13
531	push	%r14
532	push	%r15
533	pushfq
534	sub	\$64,%rsp
535
536	mov	120($context),%rax	# pull context->Rax
537	mov	248($context),%rbx	# pull context->Rip
538
539	lea	.Lprologue(%rip),%r10
540	cmp	%r10,%rbx		# context->Rip<.Lprologue
541	jb	.Lin_prologue
542
543	mov	152($context),%rax	# pull context->Rsp
544
545	lea	.Lepilogue(%rip),%r10
546	cmp	%r10,%rbx		# context->Rip>=.Lepilogue
547	jae	.Lin_prologue
548
549	mov	128+32(%rax),%rax	# pull saved stack pointer
550
551	mov	-8(%rax),%rbx
552	mov	-16(%rax),%rbp
553	mov	-24(%rax),%r12
554	mov	-32(%rax),%r13
555	mov	-40(%rax),%r14
556	mov	-48(%rax),%r15
557	mov	%rbx,144($context)	# restore context->Rbx
558	mov	%rbp,160($context)	# restore context->Rbp
559	mov	%r12,216($context)	# restore context->R12
560	mov	%r13,224($context)	# restore context->R13
561	mov	%r14,232($context)	# restore context->R14
562	mov	%r15,240($context)	# restore context->R15
563
564.Lin_prologue:
565	mov	8(%rax),%rdi
566	mov	16(%rax),%rsi
567	mov	%rax,152($context)	# restore context->Rsp
568	mov	%rsi,168($context)	# restore context->Rsi
569	mov	%rdi,176($context)	# restore context->Rdi
570
571	mov	40($disp),%rdi		# disp->ContextRecord
572	mov	$context,%rsi		# context
573	mov	\$154,%ecx		# sizeof(CONTEXT)
574	.long	0xa548f3fc		# cld; rep movsq
575
576	mov	$disp,%rsi
577	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
578	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
579	mov	0(%rsi),%r8		# arg3, disp->ControlPc
580	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
581	mov	40(%rsi),%r10		# disp->ContextRecord
582	lea	56(%rsi),%r11		# &disp->HandlerData
583	lea	24(%rsi),%r12		# &disp->EstablisherFrame
584	mov	%r10,32(%rsp)		# arg5
585	mov	%r11,40(%rsp)		# arg6
586	mov	%r12,48(%rsp)		# arg7
587	mov	%rcx,56(%rsp)		# arg8, (NULL)
588	call	*__imp_RtlVirtualUnwind(%rip)
589
590	mov	\$1,%eax		# ExceptionContinueSearch
591	add	\$64,%rsp
592	popfq
593	pop	%r15
594	pop	%r14
595	pop	%r13
596	pop	%r12
597	pop	%rbp
598	pop	%rbx
599	pop	%rdi
600	pop	%rsi
601	ret
602.size	se_handler,.-se_handler
603
604.section	.pdata
605.align	4
606	.rva	.LSEH_begin_$func
607	.rva	.LSEH_end_$func
608	.rva	.LSEH_info_$func
609
610.section	.xdata
611.align	8
612.LSEH_info_$func:
613	.byte	9,0,0,0
614	.rva	se_handler
615___
616}
617
618$code =~ s/\`([^\`]*)\`/eval $1/gem;
619print $code;
620close STDOUT or die "error closing STDOUT: $!";
621