1#! /usr/bin/env perl
2# Copyright 2010-2016 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 <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16#
17# March, June 2010
18#
19# The module implements "4-bit" GCM GHASH function and underlying
20# single multiplication operation in GF(2^128). "4-bit" means that
21# it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
22# function features so called "528B" variant utilizing additional
23# 256+16 bytes of per-key storage [+512 bytes shared table].
24# Performance results are for this streamed GHASH subroutine and are
25# expressed in cycles per processed byte, less is better:
26#
27#		gcc 3.4.x(*)	assembler
28#
29# P4		28.6		14.0		+100%
30# Opteron	19.3		7.7		+150%
31# Core2		17.8		8.1(**)		+120%
32# Atom		31.6		16.8		+88%
33# VIA Nano	21.8		10.1		+115%
34#
35# (*)	comparison is not completely fair, because C results are
36#	for vanilla "256B" implementation, while assembler results
37#	are for "528B";-)
38# (**)	it's mystery [to me] why Core2 result is not same as for
39#	Opteron;
40
41# May 2010
42#
43# Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
44# See ghash-x86.pl for background information and details about coding
45# techniques.
46#
47# Special thanks to David Woodhouse <dwmw2@infradead.org> for
48# providing access to a Westmere-based system on behalf of Intel
49# Open Source Technology Centre.
50
51# December 2012
52#
53# Overhaul: aggregate Karatsuba post-processing, improve ILP in
54# reduction_alg9, increase reduction aggregate factor to 4x. As for
55# the latter. ghash-x86.pl discusses that it makes lesser sense to
56# increase aggregate factor. Then why increase here? Critical path
57# consists of 3 independent pclmulqdq instructions, Karatsuba post-
58# processing and reduction. "On top" of this we lay down aggregated
59# multiplication operations, triplets of independent pclmulqdq's. As
60# issue rate for pclmulqdq is limited, it makes lesser sense to
61# aggregate more multiplications than it takes to perform remaining
62# non-multiplication operations. 2x is near-optimal coefficient for
63# contemporary Intel CPUs (therefore modest improvement coefficient),
64# but not for Bulldozer. Latter is because logical SIMD operations
65# are twice as slow in comparison to Intel, so that critical path is
66# longer. A CPU with higher pclmulqdq issue rate would also benefit
67# from higher aggregate factor...
68#
69# Westmere	1.78(+13%)
70# Sandy Bridge	1.80(+8%)
71# Ivy Bridge	1.80(+7%)
72# Haswell	0.55(+93%) (if system doesn't support AVX)
73# Broadwell	0.45(+110%)(if system doesn't support AVX)
74# Skylake	0.44(+110%)(if system doesn't support AVX)
75# Bulldozer	1.49(+27%)
76# Silvermont	2.88(+13%)
77# Knights L	2.12(-)    (if system doesn't support AVX)
78# Goldmont	1.08(+24%)
79
80# March 2013
81#
82# ... 8x aggregate factor AVX code path is using reduction algorithm
83# suggested by Shay Gueron[1]. Even though contemporary AVX-capable
84# CPUs such as Sandy and Ivy Bridge can execute it, the code performs
85# sub-optimally in comparison to above mentioned version. But thanks
86# to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
87# it performs in 0.41 cycles per byte on Haswell processor, in
88# 0.29 on Broadwell, and in 0.36 on Skylake.
89#
90# Knights Landing achieves 1.09 cpb.
91#
92# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
93
94$flavour = shift;
95$output  = shift;
96if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
97
98$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
99
100$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
101( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
102( $xlate="${dir}../../../perlasm/x86_64-xlate.pl" and -f $xlate) or
103die "can't locate x86_64-xlate.pl";
104
105# See the notes about |$avx| in aesni-gcm-x86_64.pl; otherwise tags will be
106# computed incorrectly.
107#
108# In upstream, this is controlled by shelling out to the compiler to check
109# versions, but BoringSSL is intended to be used with pre-generated perlasm
110# output, so this isn't useful anyway.
111$avx = 1;
112
113open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
114*STDOUT=*OUT;
115
116$do4xaggr=1;
117
118# common register layout
119$nlo="%rax";
120$nhi="%rbx";
121$Zlo="%r8";
122$Zhi="%r9";
123$tmp="%r10";
124$rem_4bit = "%r11";
125
126$Xi="%rdi";
127$Htbl="%rsi";
128
129# per-function register layout
130$cnt="%rcx";
131$rem="%rdx";
132
133sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/	or
134			$r =~ s/%[er]([sd]i)/%\1l/	or
135			$r =~ s/%[er](bp)/%\1l/		or
136			$r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
137
138sub AUTOLOAD()		# thunk [simplified] 32-bit style perlasm
139{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
140  my $arg = pop;
141    $arg = "\$$arg" if ($arg*1 eq $arg);
142    $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
143}
144
145{ my $N;
146  sub loop() {
147  my $inp = shift;
148
149	$N++;
150$code.=<<___;
151	xor	$nlo,$nlo
152	xor	$nhi,$nhi
153	mov	`&LB("$Zlo")`,`&LB("$nlo")`
154	mov	`&LB("$Zlo")`,`&LB("$nhi")`
155	shl	\$4,`&LB("$nlo")`
156	mov	\$14,$cnt
157	mov	8($Htbl,$nlo),$Zlo
158	mov	($Htbl,$nlo),$Zhi
159	and	\$0xf0,`&LB("$nhi")`
160	mov	$Zlo,$rem
161	jmp	.Loop$N
162
163.align	16
164.Loop$N:
165	shr	\$4,$Zlo
166	and	\$0xf,$rem
167	mov	$Zhi,$tmp
168	mov	($inp,$cnt),`&LB("$nlo")`
169	shr	\$4,$Zhi
170	xor	8($Htbl,$nhi),$Zlo
171	shl	\$60,$tmp
172	xor	($Htbl,$nhi),$Zhi
173	mov	`&LB("$nlo")`,`&LB("$nhi")`
174	xor	($rem_4bit,$rem,8),$Zhi
175	mov	$Zlo,$rem
176	shl	\$4,`&LB("$nlo")`
177	xor	$tmp,$Zlo
178	dec	$cnt
179	js	.Lbreak$N
180
181	shr	\$4,$Zlo
182	and	\$0xf,$rem
183	mov	$Zhi,$tmp
184	shr	\$4,$Zhi
185	xor	8($Htbl,$nlo),$Zlo
186	shl	\$60,$tmp
187	xor	($Htbl,$nlo),$Zhi
188	and	\$0xf0,`&LB("$nhi")`
189	xor	($rem_4bit,$rem,8),$Zhi
190	mov	$Zlo,$rem
191	xor	$tmp,$Zlo
192	jmp	.Loop$N
193
194.align	16
195.Lbreak$N:
196	shr	\$4,$Zlo
197	and	\$0xf,$rem
198	mov	$Zhi,$tmp
199	shr	\$4,$Zhi
200	xor	8($Htbl,$nlo),$Zlo
201	shl	\$60,$tmp
202	xor	($Htbl,$nlo),$Zhi
203	and	\$0xf0,`&LB("$nhi")`
204	xor	($rem_4bit,$rem,8),$Zhi
205	mov	$Zlo,$rem
206	xor	$tmp,$Zlo
207
208	shr	\$4,$Zlo
209	and	\$0xf,$rem
210	mov	$Zhi,$tmp
211	shr	\$4,$Zhi
212	xor	8($Htbl,$nhi),$Zlo
213	shl	\$60,$tmp
214	xor	($Htbl,$nhi),$Zhi
215	xor	$tmp,$Zlo
216	xor	($rem_4bit,$rem,8),$Zhi
217
218	bswap	$Zlo
219	bswap	$Zhi
220___
221}}
222
223$code=<<___;
224.text
225.extern	OPENSSL_ia32cap_P
226
227.globl	gcm_gmult_4bit
228.type	gcm_gmult_4bit,\@function,2
229.align	16
230gcm_gmult_4bit:
231	push	%rbx
232	push	%rbp		# %rbp and others are pushed exclusively in
233	push	%r12		# order to reuse Win64 exception handler...
234	push	%r13
235	push	%r14
236	push	%r15
237	sub	\$280,%rsp
238.Lgmult_prologue:
239
240	movzb	15($Xi),$Zlo
241	lea	.Lrem_4bit(%rip),$rem_4bit
242___
243	&loop	($Xi);
244$code.=<<___;
245	mov	$Zlo,8($Xi)
246	mov	$Zhi,($Xi)
247
248	lea	280+48(%rsp),%rsi
249	mov	-8(%rsi),%rbx
250	lea	(%rsi),%rsp
251.Lgmult_epilogue:
252	ret
253.size	gcm_gmult_4bit,.-gcm_gmult_4bit
254___
255
256# per-function register layout
257$inp="%rdx";
258$len="%rcx";
259$rem_8bit=$rem_4bit;
260
261$code.=<<___;
262.globl	gcm_ghash_4bit
263.type	gcm_ghash_4bit,\@function,4
264.align	16
265gcm_ghash_4bit:
266	push	%rbx
267	push	%rbp
268	push	%r12
269	push	%r13
270	push	%r14
271	push	%r15
272	sub	\$280,%rsp
273.Lghash_prologue:
274	mov	$inp,%r14		# reassign couple of args
275	mov	$len,%r15
276___
277{ my $inp="%r14";
278  my $dat="%edx";
279  my $len="%r15";
280  my @nhi=("%ebx","%ecx");
281  my @rem=("%r12","%r13");
282  my $Hshr4="%rbp";
283
284	&sub	($Htbl,-128);		# size optimization
285	&lea	($Hshr4,"16+128(%rsp)");
286	{ my @lo =($nlo,$nhi);
287          my @hi =($Zlo,$Zhi);
288
289	  &xor	($dat,$dat);
290	  for ($i=0,$j=-2;$i<18;$i++,$j++) {
291	    &mov	("$j(%rsp)",&LB($dat))		if ($i>1);
292	    &or		($lo[0],$tmp)			if ($i>1);
293	    &mov	(&LB($dat),&LB($lo[1]))		if ($i>0 && $i<17);
294	    &shr	($lo[1],4)			if ($i>0 && $i<17);
295	    &mov	($tmp,$hi[1])			if ($i>0 && $i<17);
296	    &shr	($hi[1],4)			if ($i>0 && $i<17);
297	    &mov	("8*$j($Hshr4)",$hi[0])		if ($i>1);
298	    &mov	($hi[0],"16*$i+0-128($Htbl)")	if ($i<16);
299	    &shl	(&LB($dat),4)			if ($i>0 && $i<17);
300	    &mov	("8*$j-128($Hshr4)",$lo[0])	if ($i>1);
301	    &mov	($lo[0],"16*$i+8-128($Htbl)")	if ($i<16);
302	    &shl	($tmp,60)			if ($i>0 && $i<17);
303
304	    push	(@lo,shift(@lo));
305	    push	(@hi,shift(@hi));
306	  }
307	}
308	&add	($Htbl,-128);
309	&mov	($Zlo,"8($Xi)");
310	&mov	($Zhi,"0($Xi)");
311	&add	($len,$inp);		# pointer to the end of data
312	&lea	($rem_8bit,".Lrem_8bit(%rip)");
313	&jmp	(".Louter_loop");
314
315$code.=".align	16\n.Louter_loop:\n";
316	&xor	($Zhi,"($inp)");
317	&mov	("%rdx","8($inp)");
318	&lea	($inp,"16($inp)");
319	&xor	("%rdx",$Zlo);
320	&mov	("($Xi)",$Zhi);
321	&mov	("8($Xi)","%rdx");
322	&shr	("%rdx",32);
323
324	&xor	($nlo,$nlo);
325	&rol	($dat,8);
326	&mov	(&LB($nlo),&LB($dat));
327	&movz	($nhi[0],&LB($dat));
328	&shl	(&LB($nlo),4);
329	&shr	($nhi[0],4);
330
331	for ($j=11,$i=0;$i<15;$i++) {
332	    &rol	($dat,8);
333	    &xor	($Zlo,"8($Htbl,$nlo)")			if ($i>0);
334	    &xor	($Zhi,"($Htbl,$nlo)")			if ($i>0);
335	    &mov	($Zlo,"8($Htbl,$nlo)")			if ($i==0);
336	    &mov	($Zhi,"($Htbl,$nlo)")			if ($i==0);
337
338	    &mov	(&LB($nlo),&LB($dat));
339	    &xor	($Zlo,$tmp)				if ($i>0);
340	    &movzw	($rem[1],"($rem_8bit,$rem[1],2)")	if ($i>0);
341
342	    &movz	($nhi[1],&LB($dat));
343	    &shl	(&LB($nlo),4);
344	    &movzb	($rem[0],"(%rsp,$nhi[0])");
345
346	    &shr	($nhi[1],4)				if ($i<14);
347	    &and	($nhi[1],0xf0)				if ($i==14);
348	    &shl	($rem[1],48)				if ($i>0);
349	    &xor	($rem[0],$Zlo);
350
351	    &mov	($tmp,$Zhi);
352	    &xor	($Zhi,$rem[1])				if ($i>0);
353	    &shr	($Zlo,8);
354
355	    &movz	($rem[0],&LB($rem[0]));
356	    &mov	($dat,"$j($Xi)")			if (--$j%4==0);
357	    &shr	($Zhi,8);
358
359	    &xor	($Zlo,"-128($Hshr4,$nhi[0],8)");
360	    &shl	($tmp,56);
361	    &xor	($Zhi,"($Hshr4,$nhi[0],8)");
362
363	    unshift	(@nhi,pop(@nhi));		# "rotate" registers
364	    unshift	(@rem,pop(@rem));
365	}
366	&movzw	($rem[1],"($rem_8bit,$rem[1],2)");
367	&xor	($Zlo,"8($Htbl,$nlo)");
368	&xor	($Zhi,"($Htbl,$nlo)");
369
370	&shl	($rem[1],48);
371	&xor	($Zlo,$tmp);
372
373	&xor	($Zhi,$rem[1]);
374	&movz	($rem[0],&LB($Zlo));
375	&shr	($Zlo,4);
376
377	&mov	($tmp,$Zhi);
378	&shl	(&LB($rem[0]),4);
379	&shr	($Zhi,4);
380
381	&xor	($Zlo,"8($Htbl,$nhi[0])");
382	&movzw	($rem[0],"($rem_8bit,$rem[0],2)");
383	&shl	($tmp,60);
384
385	&xor	($Zhi,"($Htbl,$nhi[0])");
386	&xor	($Zlo,$tmp);
387	&shl	($rem[0],48);
388
389	&bswap	($Zlo);
390	&xor	($Zhi,$rem[0]);
391
392	&bswap	($Zhi);
393	&cmp	($inp,$len);
394	&jb	(".Louter_loop");
395}
396$code.=<<___;
397	mov	$Zlo,8($Xi)
398	mov	$Zhi,($Xi)
399
400	lea	280+48(%rsp),%rsi
401	mov	-48(%rsi),%r15
402	mov	-40(%rsi),%r14
403	mov	-32(%rsi),%r13
404	mov	-24(%rsi),%r12
405	mov	-16(%rsi),%rbp
406	mov	-8(%rsi),%rbx
407	lea	0(%rsi),%rsp
408.Lghash_epilogue:
409	ret
410.size	gcm_ghash_4bit,.-gcm_ghash_4bit
411___
412
413######################################################################
414# PCLMULQDQ version.
415
416@_4args=$win64?	("%rcx","%rdx","%r8", "%r9") :	# Win64 order
417		("%rdi","%rsi","%rdx","%rcx");	# Unix order
418
419($Xi,$Xhi)=("%xmm0","%xmm1");	$Hkey="%xmm2";
420($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
421
422sub clmul64x64_T2 {	# minimal register pressure
423my ($Xhi,$Xi,$Hkey,$HK)=@_;
424
425if (!defined($HK)) {	$HK = $T2;
426$code.=<<___;
427	movdqa		$Xi,$Xhi		#
428	pshufd		\$0b01001110,$Xi,$T1
429	pshufd		\$0b01001110,$Hkey,$T2
430	pxor		$Xi,$T1			#
431	pxor		$Hkey,$T2
432___
433} else {
434$code.=<<___;
435	movdqa		$Xi,$Xhi		#
436	pshufd		\$0b01001110,$Xi,$T1
437	pxor		$Xi,$T1			#
438___
439}
440$code.=<<___;
441	pclmulqdq	\$0x00,$Hkey,$Xi	#######
442	pclmulqdq	\$0x11,$Hkey,$Xhi	#######
443	pclmulqdq	\$0x00,$HK,$T1		#######
444	pxor		$Xi,$T1			#
445	pxor		$Xhi,$T1		#
446
447	movdqa		$T1,$T2			#
448	psrldq		\$8,$T1
449	pslldq		\$8,$T2			#
450	pxor		$T1,$Xhi
451	pxor		$T2,$Xi			#
452___
453}
454
455sub reduction_alg9 {	# 17/11 times faster than Intel version
456my ($Xhi,$Xi) = @_;
457
458$code.=<<___;
459	# 1st phase
460	movdqa		$Xi,$T2			#
461	movdqa		$Xi,$T1
462	psllq		\$5,$Xi
463	pxor		$Xi,$T1			#
464	psllq		\$1,$Xi
465	pxor		$T1,$Xi			#
466	psllq		\$57,$Xi		#
467	movdqa		$Xi,$T1			#
468	pslldq		\$8,$Xi
469	psrldq		\$8,$T1			#
470	pxor		$T2,$Xi
471	pxor		$T1,$Xhi		#
472
473	# 2nd phase
474	movdqa		$Xi,$T2
475	psrlq		\$1,$Xi
476	pxor		$T2,$Xhi		#
477	pxor		$Xi,$T2
478	psrlq		\$5,$Xi
479	pxor		$T2,$Xi			#
480	psrlq		\$1,$Xi			#
481	pxor		$Xhi,$Xi		#
482___
483}
484
485{ my ($Htbl,$Xip)=@_4args;
486  my $HK="%xmm6";
487
488$code.=<<___;
489.globl	gcm_init_clmul
490.type	gcm_init_clmul,\@abi-omnipotent
491.align	16
492gcm_init_clmul:
493.L_init_clmul:
494___
495$code.=<<___ if ($win64);
496.LSEH_begin_gcm_init_clmul:
497	# I can't trust assembler to use specific encoding:-(
498	.byte	0x48,0x83,0xec,0x18		#sub	$0x18,%rsp
499	.byte	0x0f,0x29,0x34,0x24		#movaps	%xmm6,(%rsp)
500___
501$code.=<<___;
502	movdqu		($Xip),$Hkey
503	pshufd		\$0b01001110,$Hkey,$Hkey	# dword swap
504
505	# <<1 twist
506	pshufd		\$0b11111111,$Hkey,$T2	# broadcast uppermost dword
507	movdqa		$Hkey,$T1
508	psllq		\$1,$Hkey
509	pxor		$T3,$T3			#
510	psrlq		\$63,$T1
511	pcmpgtd		$T2,$T3			# broadcast carry bit
512	pslldq		\$8,$T1
513	por		$T1,$Hkey		# H<<=1
514
515	# magic reduction
516	pand		.L0x1c2_polynomial(%rip),$T3
517	pxor		$T3,$Hkey		# if(carry) H^=0x1c2_polynomial
518
519	# calculate H^2
520	pshufd		\$0b01001110,$Hkey,$HK
521	movdqa		$Hkey,$Xi
522	pxor		$Hkey,$HK
523___
524	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);
525	&reduction_alg9	($Xhi,$Xi);
526$code.=<<___;
527	pshufd		\$0b01001110,$Hkey,$T1
528	pshufd		\$0b01001110,$Xi,$T2
529	pxor		$Hkey,$T1		# Karatsuba pre-processing
530	movdqu		$Hkey,0x00($Htbl)	# save H
531	pxor		$Xi,$T2			# Karatsuba pre-processing
532	movdqu		$Xi,0x10($Htbl)		# save H^2
533	palignr		\$8,$T1,$T2		# low part is H.lo^H.hi...
534	movdqu		$T2,0x20($Htbl)		# save Karatsuba "salt"
535___
536if ($do4xaggr) {
537	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H^3
538	&reduction_alg9	($Xhi,$Xi);
539$code.=<<___;
540	movdqa		$Xi,$T3
541___
542	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H^4
543	&reduction_alg9	($Xhi,$Xi);
544$code.=<<___;
545	pshufd		\$0b01001110,$T3,$T1
546	pshufd		\$0b01001110,$Xi,$T2
547	pxor		$T3,$T1			# Karatsuba pre-processing
548	movdqu		$T3,0x30($Htbl)		# save H^3
549	pxor		$Xi,$T2			# Karatsuba pre-processing
550	movdqu		$Xi,0x40($Htbl)		# save H^4
551	palignr		\$8,$T1,$T2		# low part is H^3.lo^H^3.hi...
552	movdqu		$T2,0x50($Htbl)		# save Karatsuba "salt"
553___
554}
555$code.=<<___ if ($win64);
556	movaps	(%rsp),%xmm6
557	lea	0x18(%rsp),%rsp
558.LSEH_end_gcm_init_clmul:
559___
560$code.=<<___;
561	ret
562.size	gcm_init_clmul,.-gcm_init_clmul
563___
564}
565
566{ my ($Xip,$Htbl)=@_4args;
567
568$code.=<<___;
569.globl	gcm_gmult_clmul
570.type	gcm_gmult_clmul,\@abi-omnipotent
571.align	16
572gcm_gmult_clmul:
573.L_gmult_clmul:
574	movdqu		($Xip),$Xi
575	movdqa		.Lbswap_mask(%rip),$T3
576	movdqu		($Htbl),$Hkey
577	movdqu		0x20($Htbl),$T2
578	pshufb		$T3,$Xi
579___
580	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$T2);
581$code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
582	# experimental alternative. special thing about is that there
583	# no dependency between the two multiplications...
584	mov		\$`0xE1<<1`,%eax
585	mov		\$0xA040608020C0E000,%r10	# ((7..0)��0xE0)&0xff
586	mov		\$0x07,%r11d
587	movq		%rax,$T1
588	movq		%r10,$T2
589	movq		%r11,$T3		# borrow $T3
590	pand		$Xi,$T3
591	pshufb		$T3,$T2			# ($Xi&7)��0xE0
592	movq		%rax,$T3
593	pclmulqdq	\$0x00,$Xi,$T1		# ��(0xE1<<1)
594	pxor		$Xi,$T2
595	pslldq		\$15,$T2
596	paddd		$T2,$T2			# <<(64+56+1)
597	pxor		$T2,$Xi
598	pclmulqdq	\$0x01,$T3,$Xi
599	movdqa		.Lbswap_mask(%rip),$T3	# reload $T3
600	psrldq		\$1,$T1
601	pxor		$T1,$Xhi
602	pslldq		\$7,$Xi
603	pxor		$Xhi,$Xi
604___
605$code.=<<___;
606	pshufb		$T3,$Xi
607	movdqu		$Xi,($Xip)
608	ret
609.size	gcm_gmult_clmul,.-gcm_gmult_clmul
610___
611}
612
613{ my ($Xip,$Htbl,$inp,$len)=@_4args;
614  my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
615  my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
616
617$code.=<<___;
618.globl	gcm_ghash_clmul
619.type	gcm_ghash_clmul,\@abi-omnipotent
620.align	32
621gcm_ghash_clmul:
622.L_ghash_clmul:
623___
624$code.=<<___ if ($win64);
625	lea	-0x88(%rsp),%rax
626.LSEH_begin_gcm_ghash_clmul:
627	# I can't trust assembler to use specific encoding:-(
628	.byte	0x48,0x8d,0x60,0xe0		#lea	-0x20(%rax),%rsp
629	.byte	0x0f,0x29,0x70,0xe0		#movaps	%xmm6,-0x20(%rax)
630	.byte	0x0f,0x29,0x78,0xf0		#movaps	%xmm7,-0x10(%rax)
631	.byte	0x44,0x0f,0x29,0x00		#movaps	%xmm8,0(%rax)
632	.byte	0x44,0x0f,0x29,0x48,0x10	#movaps	%xmm9,0x10(%rax)
633	.byte	0x44,0x0f,0x29,0x50,0x20	#movaps	%xmm10,0x20(%rax)
634	.byte	0x44,0x0f,0x29,0x58,0x30	#movaps	%xmm11,0x30(%rax)
635	.byte	0x44,0x0f,0x29,0x60,0x40	#movaps	%xmm12,0x40(%rax)
636	.byte	0x44,0x0f,0x29,0x68,0x50	#movaps	%xmm13,0x50(%rax)
637	.byte	0x44,0x0f,0x29,0x70,0x60	#movaps	%xmm14,0x60(%rax)
638	.byte	0x44,0x0f,0x29,0x78,0x70	#movaps	%xmm15,0x70(%rax)
639___
640$code.=<<___;
641	movdqa		.Lbswap_mask(%rip),$T3
642
643	movdqu		($Xip),$Xi
644	movdqu		($Htbl),$Hkey
645	movdqu		0x20($Htbl),$HK
646	pshufb		$T3,$Xi
647
648	sub		\$0x10,$len
649	jz		.Lodd_tail
650
651	movdqu		0x10($Htbl),$Hkey2
652___
653if ($do4xaggr) {
654my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
655
656$code.=<<___;
657	leaq		OPENSSL_ia32cap_P(%rip),%rax
658	mov		4(%rax),%eax
659	cmp		\$0x30,$len
660	jb		.Lskip4x
661
662	and		\$`1<<26|1<<22`,%eax	# isolate MOVBE+XSAVE
663	cmp		\$`1<<22`,%eax		# check for MOVBE without XSAVE
664	je		.Lskip4x
665
666	sub		\$0x30,$len
667	mov		\$0xA040608020C0E000,%rax	# ((7..0)��0xE0)&0xff
668	movdqu		0x30($Htbl),$Hkey3
669	movdqu		0x40($Htbl),$Hkey4
670
671	#######
672	# Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
673	#
674	movdqu		0x30($inp),$Xln
675	 movdqu		0x20($inp),$Xl
676	pshufb		$T3,$Xln
677	 pshufb		$T3,$Xl
678	movdqa		$Xln,$Xhn
679	pshufd		\$0b01001110,$Xln,$Xmn
680	pxor		$Xln,$Xmn
681	pclmulqdq	\$0x00,$Hkey,$Xln
682	pclmulqdq	\$0x11,$Hkey,$Xhn
683	pclmulqdq	\$0x00,$HK,$Xmn
684
685	movdqa		$Xl,$Xh
686	pshufd		\$0b01001110,$Xl,$Xm
687	pxor		$Xl,$Xm
688	pclmulqdq	\$0x00,$Hkey2,$Xl
689	pclmulqdq	\$0x11,$Hkey2,$Xh
690	pclmulqdq	\$0x10,$HK,$Xm
691	xorps		$Xl,$Xln
692	xorps		$Xh,$Xhn
693	movups		0x50($Htbl),$HK
694	xorps		$Xm,$Xmn
695
696	movdqu		0x10($inp),$Xl
697	 movdqu		0($inp),$T1
698	pshufb		$T3,$Xl
699	 pshufb		$T3,$T1
700	movdqa		$Xl,$Xh
701	pshufd		\$0b01001110,$Xl,$Xm
702	 pxor		$T1,$Xi
703	pxor		$Xl,$Xm
704	pclmulqdq	\$0x00,$Hkey3,$Xl
705	 movdqa		$Xi,$Xhi
706	 pshufd		\$0b01001110,$Xi,$T1
707	 pxor		$Xi,$T1
708	pclmulqdq	\$0x11,$Hkey3,$Xh
709	pclmulqdq	\$0x00,$HK,$Xm
710	xorps		$Xl,$Xln
711	xorps		$Xh,$Xhn
712
713	lea	0x40($inp),$inp
714	sub	\$0x40,$len
715	jc	.Ltail4x
716
717	jmp	.Lmod4_loop
718.align	32
719.Lmod4_loop:
720	pclmulqdq	\$0x00,$Hkey4,$Xi
721	xorps		$Xm,$Xmn
722	 movdqu		0x30($inp),$Xl
723	 pshufb		$T3,$Xl
724	pclmulqdq	\$0x11,$Hkey4,$Xhi
725	xorps		$Xln,$Xi
726	 movdqu		0x20($inp),$Xln
727	 movdqa		$Xl,$Xh
728	pclmulqdq	\$0x10,$HK,$T1
729	 pshufd		\$0b01001110,$Xl,$Xm
730	xorps		$Xhn,$Xhi
731	 pxor		$Xl,$Xm
732	 pshufb		$T3,$Xln
733	movups		0x20($Htbl),$HK
734	xorps		$Xmn,$T1
735	 pclmulqdq	\$0x00,$Hkey,$Xl
736	 pshufd		\$0b01001110,$Xln,$Xmn
737
738	pxor		$Xi,$T1			# aggregated Karatsuba post-processing
739	 movdqa		$Xln,$Xhn
740	pxor		$Xhi,$T1		#
741	 pxor		$Xln,$Xmn
742	movdqa		$T1,$T2			#
743	 pclmulqdq	\$0x11,$Hkey,$Xh
744	pslldq		\$8,$T1
745	psrldq		\$8,$T2			#
746	pxor		$T1,$Xi
747	movdqa		.L7_mask(%rip),$T1
748	pxor		$T2,$Xhi		#
749	movq		%rax,$T2
750
751	pand		$Xi,$T1			# 1st phase
752	pshufb		$T1,$T2			#
753	pxor		$Xi,$T2			#
754	 pclmulqdq	\$0x00,$HK,$Xm
755	psllq		\$57,$T2		#
756	movdqa		$T2,$T1			#
757	pslldq		\$8,$T2
758	 pclmulqdq	\$0x00,$Hkey2,$Xln
759	psrldq		\$8,$T1			#
760	pxor		$T2,$Xi
761	pxor		$T1,$Xhi		#
762	movdqu		0($inp),$T1
763
764	movdqa		$Xi,$T2			# 2nd phase
765	psrlq		\$1,$Xi
766	 pclmulqdq	\$0x11,$Hkey2,$Xhn
767	 xorps		$Xl,$Xln
768	 movdqu		0x10($inp),$Xl
769	 pshufb		$T3,$Xl
770	 pclmulqdq	\$0x10,$HK,$Xmn
771	 xorps		$Xh,$Xhn
772	 movups		0x50($Htbl),$HK
773	pshufb		$T3,$T1
774	pxor		$T2,$Xhi		#
775	pxor		$Xi,$T2
776	psrlq		\$5,$Xi
777
778	 movdqa		$Xl,$Xh
779	 pxor		$Xm,$Xmn
780	 pshufd		\$0b01001110,$Xl,$Xm
781	pxor		$T2,$Xi			#
782	pxor		$T1,$Xhi
783	 pxor		$Xl,$Xm
784	 pclmulqdq	\$0x00,$Hkey3,$Xl
785	psrlq		\$1,$Xi			#
786	pxor		$Xhi,$Xi		#
787	movdqa		$Xi,$Xhi
788	 pclmulqdq	\$0x11,$Hkey3,$Xh
789	 xorps		$Xl,$Xln
790	pshufd		\$0b01001110,$Xi,$T1
791	pxor		$Xi,$T1
792
793	 pclmulqdq	\$0x00,$HK,$Xm
794	 xorps		$Xh,$Xhn
795
796	lea	0x40($inp),$inp
797	sub	\$0x40,$len
798	jnc	.Lmod4_loop
799
800.Ltail4x:
801	pclmulqdq	\$0x00,$Hkey4,$Xi
802	pclmulqdq	\$0x11,$Hkey4,$Xhi
803	pclmulqdq	\$0x10,$HK,$T1
804	xorps		$Xm,$Xmn
805	xorps		$Xln,$Xi
806	xorps		$Xhn,$Xhi
807	pxor		$Xi,$Xhi		# aggregated Karatsuba post-processing
808	pxor		$Xmn,$T1
809
810	pxor		$Xhi,$T1		#
811	pxor		$Xi,$Xhi
812
813	movdqa		$T1,$T2			#
814	psrldq		\$8,$T1
815	pslldq		\$8,$T2			#
816	pxor		$T1,$Xhi
817	pxor		$T2,$Xi			#
818___
819	&reduction_alg9($Xhi,$Xi);
820$code.=<<___;
821	add	\$0x40,$len
822	jz	.Ldone
823	movdqu	0x20($Htbl),$HK
824	sub	\$0x10,$len
825	jz	.Lodd_tail
826.Lskip4x:
827___
828}
829$code.=<<___;
830	#######
831	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
832	#	[(H*Ii+1) + (H*Xi+1)] mod P =
833	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
834	#
835	movdqu		($inp),$T1		# Ii
836	movdqu		16($inp),$Xln		# Ii+1
837	pshufb		$T3,$T1
838	pshufb		$T3,$Xln
839	pxor		$T1,$Xi			# Ii+Xi
840
841	movdqa		$Xln,$Xhn
842	pshufd		\$0b01001110,$Xln,$Xmn
843	pxor		$Xln,$Xmn
844	pclmulqdq	\$0x00,$Hkey,$Xln
845	pclmulqdq	\$0x11,$Hkey,$Xhn
846	pclmulqdq	\$0x00,$HK,$Xmn
847
848	lea		32($inp),$inp		# i+=2
849	nop
850	sub		\$0x20,$len
851	jbe		.Leven_tail
852	nop
853	jmp		.Lmod_loop
854
855.align	32
856.Lmod_loop:
857	movdqa		$Xi,$Xhi
858	movdqa		$Xmn,$T1
859	pshufd		\$0b01001110,$Xi,$Xmn	#
860	pxor		$Xi,$Xmn		#
861
862	pclmulqdq	\$0x00,$Hkey2,$Xi
863	pclmulqdq	\$0x11,$Hkey2,$Xhi
864	pclmulqdq	\$0x10,$HK,$Xmn
865
866	pxor		$Xln,$Xi		# (H*Ii+1) + H^2*(Ii+Xi)
867	pxor		$Xhn,$Xhi
868	  movdqu	($inp),$T2		# Ii
869	pxor		$Xi,$T1			# aggregated Karatsuba post-processing
870	  pshufb	$T3,$T2
871	  movdqu	16($inp),$Xln		# Ii+1
872
873	pxor		$Xhi,$T1
874	  pxor		$T2,$Xhi		# "Ii+Xi", consume early
875	pxor		$T1,$Xmn
876	 pshufb		$T3,$Xln
877	movdqa		$Xmn,$T1		#
878	psrldq		\$8,$T1
879	pslldq		\$8,$Xmn		#
880	pxor		$T1,$Xhi
881	pxor		$Xmn,$Xi		#
882
883	movdqa		$Xln,$Xhn		#
884
885	  movdqa	$Xi,$T2			# 1st phase
886	  movdqa	$Xi,$T1
887	  psllq		\$5,$Xi
888	  pxor		$Xi,$T1			#
889	pclmulqdq	\$0x00,$Hkey,$Xln	#######
890	  psllq		\$1,$Xi
891	  pxor		$T1,$Xi			#
892	  psllq		\$57,$Xi		#
893	  movdqa	$Xi,$T1			#
894	  pslldq	\$8,$Xi
895	  psrldq	\$8,$T1			#
896	  pxor		$T2,$Xi
897	pshufd		\$0b01001110,$Xhn,$Xmn
898	  pxor		$T1,$Xhi		#
899	pxor		$Xhn,$Xmn		#
900
901	  movdqa	$Xi,$T2			# 2nd phase
902	  psrlq		\$1,$Xi
903	pclmulqdq	\$0x11,$Hkey,$Xhn	#######
904	  pxor		$T2,$Xhi		#
905	  pxor		$Xi,$T2
906	  psrlq		\$5,$Xi
907	  pxor		$T2,$Xi			#
908	lea		32($inp),$inp
909	  psrlq		\$1,$Xi			#
910	pclmulqdq	\$0x00,$HK,$Xmn		#######
911	  pxor		$Xhi,$Xi		#
912
913	sub		\$0x20,$len
914	ja		.Lmod_loop
915
916.Leven_tail:
917	 movdqa		$Xi,$Xhi
918	 movdqa		$Xmn,$T1
919	 pshufd		\$0b01001110,$Xi,$Xmn	#
920	 pxor		$Xi,$Xmn		#
921
922	pclmulqdq	\$0x00,$Hkey2,$Xi
923	pclmulqdq	\$0x11,$Hkey2,$Xhi
924	pclmulqdq	\$0x10,$HK,$Xmn
925
926	pxor		$Xln,$Xi		# (H*Ii+1) + H^2*(Ii+Xi)
927	pxor		$Xhn,$Xhi
928	pxor		$Xi,$T1
929	pxor		$Xhi,$T1
930	pxor		$T1,$Xmn
931	movdqa		$Xmn,$T1		#
932	psrldq		\$8,$T1
933	pslldq		\$8,$Xmn		#
934	pxor		$T1,$Xhi
935	pxor		$Xmn,$Xi		#
936___
937	&reduction_alg9	($Xhi,$Xi);
938$code.=<<___;
939	test		$len,$len
940	jnz		.Ldone
941
942.Lodd_tail:
943	movdqu		($inp),$T1		# Ii
944	pshufb		$T3,$T1
945	pxor		$T1,$Xi			# Ii+Xi
946___
947	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H*(Ii+Xi)
948	&reduction_alg9	($Xhi,$Xi);
949$code.=<<___;
950.Ldone:
951	pshufb		$T3,$Xi
952	movdqu		$Xi,($Xip)
953___
954$code.=<<___ if ($win64);
955	movaps	(%rsp),%xmm6
956	movaps	0x10(%rsp),%xmm7
957	movaps	0x20(%rsp),%xmm8
958	movaps	0x30(%rsp),%xmm9
959	movaps	0x40(%rsp),%xmm10
960	movaps	0x50(%rsp),%xmm11
961	movaps	0x60(%rsp),%xmm12
962	movaps	0x70(%rsp),%xmm13
963	movaps	0x80(%rsp),%xmm14
964	movaps	0x90(%rsp),%xmm15
965	lea	0xa8(%rsp),%rsp
966.LSEH_end_gcm_ghash_clmul:
967___
968$code.=<<___;
969	ret
970.size	gcm_ghash_clmul,.-gcm_ghash_clmul
971___
972}
973
974$code.=<<___;
975.globl	gcm_init_avx
976.type	gcm_init_avx,\@abi-omnipotent
977.align	32
978gcm_init_avx:
979___
980if ($avx) {
981my ($Htbl,$Xip)=@_4args;
982my $HK="%xmm6";
983
984$code.=<<___ if ($win64);
985.LSEH_begin_gcm_init_avx:
986	# I can't trust assembler to use specific encoding:-(
987	.byte	0x48,0x83,0xec,0x18		#sub	$0x18,%rsp
988	.byte	0x0f,0x29,0x34,0x24		#movaps	%xmm6,(%rsp)
989___
990$code.=<<___;
991	vzeroupper
992
993	vmovdqu		($Xip),$Hkey
994	vpshufd		\$0b01001110,$Hkey,$Hkey	# dword swap
995
996	# <<1 twist
997	vpshufd		\$0b11111111,$Hkey,$T2	# broadcast uppermost dword
998	vpsrlq		\$63,$Hkey,$T1
999	vpsllq		\$1,$Hkey,$Hkey
1000	vpxor		$T3,$T3,$T3		#
1001	vpcmpgtd	$T2,$T3,$T3		# broadcast carry bit
1002	vpslldq		\$8,$T1,$T1
1003	vpor		$T1,$Hkey,$Hkey		# H<<=1
1004
1005	# magic reduction
1006	vpand		.L0x1c2_polynomial(%rip),$T3,$T3
1007	vpxor		$T3,$Hkey,$Hkey		# if(carry) H^=0x1c2_polynomial
1008
1009	vpunpckhqdq	$Hkey,$Hkey,$HK
1010	vmovdqa		$Hkey,$Xi
1011	vpxor		$Hkey,$HK,$HK
1012	mov		\$4,%r10		# up to H^8
1013	jmp		.Linit_start_avx
1014___
1015
1016sub clmul64x64_avx {
1017my ($Xhi,$Xi,$Hkey,$HK)=@_;
1018
1019if (!defined($HK)) {	$HK = $T2;
1020$code.=<<___;
1021	vpunpckhqdq	$Xi,$Xi,$T1
1022	vpunpckhqdq	$Hkey,$Hkey,$T2
1023	vpxor		$Xi,$T1,$T1		#
1024	vpxor		$Hkey,$T2,$T2
1025___
1026} else {
1027$code.=<<___;
1028	vpunpckhqdq	$Xi,$Xi,$T1
1029	vpxor		$Xi,$T1,$T1		#
1030___
1031}
1032$code.=<<___;
1033	vpclmulqdq	\$0x11,$Hkey,$Xi,$Xhi	#######
1034	vpclmulqdq	\$0x00,$Hkey,$Xi,$Xi	#######
1035	vpclmulqdq	\$0x00,$HK,$T1,$T1	#######
1036	vpxor		$Xi,$Xhi,$T2		#
1037	vpxor		$T2,$T1,$T1		#
1038
1039	vpslldq		\$8,$T1,$T2		#
1040	vpsrldq		\$8,$T1,$T1
1041	vpxor		$T2,$Xi,$Xi		#
1042	vpxor		$T1,$Xhi,$Xhi
1043___
1044}
1045
1046sub reduction_avx {
1047my ($Xhi,$Xi) = @_;
1048
1049$code.=<<___;
1050	vpsllq		\$57,$Xi,$T1		# 1st phase
1051	vpsllq		\$62,$Xi,$T2
1052	vpxor		$T1,$T2,$T2		#
1053	vpsllq		\$63,$Xi,$T1
1054	vpxor		$T1,$T2,$T2		#
1055	vpslldq		\$8,$T2,$T1		#
1056	vpsrldq		\$8,$T2,$T2
1057	vpxor		$T1,$Xi,$Xi		#
1058	vpxor		$T2,$Xhi,$Xhi
1059
1060	vpsrlq		\$1,$Xi,$T2		# 2nd phase
1061	vpxor		$Xi,$Xhi,$Xhi
1062	vpxor		$T2,$Xi,$Xi		#
1063	vpsrlq		\$5,$T2,$T2
1064	vpxor		$T2,$Xi,$Xi		#
1065	vpsrlq		\$1,$Xi,$Xi		#
1066	vpxor		$Xhi,$Xi,$Xi		#
1067___
1068}
1069
1070$code.=<<___;
1071.align	32
1072.Linit_loop_avx:
1073	vpalignr	\$8,$T1,$T2,$T3		# low part is H.lo^H.hi...
1074	vmovdqu		$T3,-0x10($Htbl)	# save Karatsuba "salt"
1075___
1076	&clmul64x64_avx	($Xhi,$Xi,$Hkey,$HK);	# calculate H^3,5,7
1077	&reduction_avx	($Xhi,$Xi);
1078$code.=<<___;
1079.Linit_start_avx:
1080	vmovdqa		$Xi,$T3
1081___
1082	&clmul64x64_avx	($Xhi,$Xi,$Hkey,$HK);	# calculate H^2,4,6,8
1083	&reduction_avx	($Xhi,$Xi);
1084$code.=<<___;
1085	vpshufd		\$0b01001110,$T3,$T1
1086	vpshufd		\$0b01001110,$Xi,$T2
1087	vpxor		$T3,$T1,$T1		# Karatsuba pre-processing
1088	vmovdqu		$T3,0x00($Htbl)		# save H^1,3,5,7
1089	vpxor		$Xi,$T2,$T2		# Karatsuba pre-processing
1090	vmovdqu		$Xi,0x10($Htbl)		# save H^2,4,6,8
1091	lea		0x30($Htbl),$Htbl
1092	sub		\$1,%r10
1093	jnz		.Linit_loop_avx
1094
1095	vpalignr	\$8,$T2,$T1,$T3		# last "salt" is flipped
1096	vmovdqu		$T3,-0x10($Htbl)
1097
1098	vzeroupper
1099___
1100$code.=<<___ if ($win64);
1101	movaps	(%rsp),%xmm6
1102	lea	0x18(%rsp),%rsp
1103.LSEH_end_gcm_init_avx:
1104___
1105$code.=<<___;
1106	ret
1107.size	gcm_init_avx,.-gcm_init_avx
1108___
1109} else {
1110$code.=<<___;
1111	jmp	.L_init_clmul
1112.size	gcm_init_avx,.-gcm_init_avx
1113___
1114}
1115
1116$code.=<<___;
1117.globl	gcm_gmult_avx
1118.type	gcm_gmult_avx,\@abi-omnipotent
1119.align	32
1120gcm_gmult_avx:
1121	jmp	.L_gmult_clmul
1122.size	gcm_gmult_avx,.-gcm_gmult_avx
1123___
1124
1125$code.=<<___;
1126.globl	gcm_ghash_avx
1127.type	gcm_ghash_avx,\@abi-omnipotent
1128.align	32
1129gcm_ghash_avx:
1130___
1131if ($avx) {
1132my ($Xip,$Htbl,$inp,$len)=@_4args;
1133my ($Xlo,$Xhi,$Xmi,
1134    $Zlo,$Zhi,$Zmi,
1135    $Hkey,$HK,$T1,$T2,
1136    $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1137
1138$code.=<<___ if ($win64);
1139	lea	-0x88(%rsp),%rax
1140.LSEH_begin_gcm_ghash_avx:
1141	# I can't trust assembler to use specific encoding:-(
1142	.byte	0x48,0x8d,0x60,0xe0		#lea	-0x20(%rax),%rsp
1143	.byte	0x0f,0x29,0x70,0xe0		#movaps	%xmm6,-0x20(%rax)
1144	.byte	0x0f,0x29,0x78,0xf0		#movaps	%xmm7,-0x10(%rax)
1145	.byte	0x44,0x0f,0x29,0x00		#movaps	%xmm8,0(%rax)
1146	.byte	0x44,0x0f,0x29,0x48,0x10	#movaps	%xmm9,0x10(%rax)
1147	.byte	0x44,0x0f,0x29,0x50,0x20	#movaps	%xmm10,0x20(%rax)
1148	.byte	0x44,0x0f,0x29,0x58,0x30	#movaps	%xmm11,0x30(%rax)
1149	.byte	0x44,0x0f,0x29,0x60,0x40	#movaps	%xmm12,0x40(%rax)
1150	.byte	0x44,0x0f,0x29,0x68,0x50	#movaps	%xmm13,0x50(%rax)
1151	.byte	0x44,0x0f,0x29,0x70,0x60	#movaps	%xmm14,0x60(%rax)
1152	.byte	0x44,0x0f,0x29,0x78,0x70	#movaps	%xmm15,0x70(%rax)
1153___
1154$code.=<<___;
1155	vzeroupper
1156
1157	vmovdqu		($Xip),$Xi		# load $Xi
1158	lea		.L0x1c2_polynomial(%rip),%r10
1159	lea		0x40($Htbl),$Htbl	# size optimization
1160	vmovdqu		.Lbswap_mask(%rip),$bswap
1161	vpshufb		$bswap,$Xi,$Xi
1162	cmp		\$0x80,$len
1163	jb		.Lshort_avx
1164	sub		\$0x80,$len
1165
1166	vmovdqu		0x70($inp),$Ii		# I[7]
1167	vmovdqu		0x00-0x40($Htbl),$Hkey	# $Hkey^1
1168	vpshufb		$bswap,$Ii,$Ii
1169	vmovdqu		0x20-0x40($Htbl),$HK
1170
1171	vpunpckhqdq	$Ii,$Ii,$T2
1172	 vmovdqu	0x60($inp),$Ij		# I[6]
1173	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1174	vpxor		$Ii,$T2,$T2
1175	 vpshufb	$bswap,$Ij,$Ij
1176	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1177	 vmovdqu	0x10-0x40($Htbl),$Hkey	# $Hkey^2
1178	 vpunpckhqdq	$Ij,$Ij,$T1
1179	 vmovdqu	0x50($inp),$Ii		# I[5]
1180	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1181	 vpxor		$Ij,$T1,$T1
1182
1183	 vpshufb	$bswap,$Ii,$Ii
1184	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1185	 vpunpckhqdq	$Ii,$Ii,$T2
1186	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1187	 vmovdqu	0x30-0x40($Htbl),$Hkey	# $Hkey^3
1188	 vpxor		$Ii,$T2,$T2
1189	 vmovdqu	0x40($inp),$Ij		# I[4]
1190	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1191	 vmovdqu	0x50-0x40($Htbl),$HK
1192
1193	 vpshufb	$bswap,$Ij,$Ij
1194	vpxor		$Xlo,$Zlo,$Zlo
1195	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1196	vpxor		$Xhi,$Zhi,$Zhi
1197	 vpunpckhqdq	$Ij,$Ij,$T1
1198	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1199	 vmovdqu	0x40-0x40($Htbl),$Hkey	# $Hkey^4
1200	vpxor		$Xmi,$Zmi,$Zmi
1201	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1202	 vpxor		$Ij,$T1,$T1
1203
1204	 vmovdqu	0x30($inp),$Ii		# I[3]
1205	vpxor		$Zlo,$Xlo,$Xlo
1206	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1207	vpxor		$Zhi,$Xhi,$Xhi
1208	 vpshufb	$bswap,$Ii,$Ii
1209	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1210	 vmovdqu	0x60-0x40($Htbl),$Hkey	# $Hkey^5
1211	vpxor		$Zmi,$Xmi,$Xmi
1212	 vpunpckhqdq	$Ii,$Ii,$T2
1213	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1214	 vmovdqu	0x80-0x40($Htbl),$HK
1215	 vpxor		$Ii,$T2,$T2
1216
1217	 vmovdqu	0x20($inp),$Ij		# I[2]
1218	vpxor		$Xlo,$Zlo,$Zlo
1219	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1220	vpxor		$Xhi,$Zhi,$Zhi
1221	 vpshufb	$bswap,$Ij,$Ij
1222	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1223	 vmovdqu	0x70-0x40($Htbl),$Hkey	# $Hkey^6
1224	vpxor		$Xmi,$Zmi,$Zmi
1225	 vpunpckhqdq	$Ij,$Ij,$T1
1226	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1227	 vpxor		$Ij,$T1,$T1
1228
1229	 vmovdqu	0x10($inp),$Ii		# I[1]
1230	vpxor		$Zlo,$Xlo,$Xlo
1231	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1232	vpxor		$Zhi,$Xhi,$Xhi
1233	 vpshufb	$bswap,$Ii,$Ii
1234	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1235	 vmovdqu	0x90-0x40($Htbl),$Hkey	# $Hkey^7
1236	vpxor		$Zmi,$Xmi,$Xmi
1237	 vpunpckhqdq	$Ii,$Ii,$T2
1238	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1239	 vmovdqu	0xb0-0x40($Htbl),$HK
1240	 vpxor		$Ii,$T2,$T2
1241
1242	 vmovdqu	($inp),$Ij		# I[0]
1243	vpxor		$Xlo,$Zlo,$Zlo
1244	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1245	vpxor		$Xhi,$Zhi,$Zhi
1246	 vpshufb	$bswap,$Ij,$Ij
1247	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1248	 vmovdqu	0xa0-0x40($Htbl),$Hkey	# $Hkey^8
1249	vpxor		$Xmi,$Zmi,$Zmi
1250	vpclmulqdq	\$0x10,$HK,$T2,$Xmi
1251
1252	lea		0x80($inp),$inp
1253	cmp		\$0x80,$len
1254	jb		.Ltail_avx
1255
1256	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1257	sub		\$0x80,$len
1258	jmp		.Loop8x_avx
1259
1260.align	32
1261.Loop8x_avx:
1262	vpunpckhqdq	$Ij,$Ij,$T1
1263	 vmovdqu	0x70($inp),$Ii		# I[7]
1264	vpxor		$Xlo,$Zlo,$Zlo
1265	vpxor		$Ij,$T1,$T1
1266	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xi
1267	 vpshufb	$bswap,$Ii,$Ii
1268	vpxor		$Xhi,$Zhi,$Zhi
1269	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xo
1270	 vmovdqu	0x00-0x40($Htbl),$Hkey	# $Hkey^1
1271	 vpunpckhqdq	$Ii,$Ii,$T2
1272	vpxor		$Xmi,$Zmi,$Zmi
1273	vpclmulqdq	\$0x00,$HK,$T1,$Tred
1274	 vmovdqu	0x20-0x40($Htbl),$HK
1275	 vpxor		$Ii,$T2,$T2
1276
1277	  vmovdqu	0x60($inp),$Ij		# I[6]
1278	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1279	vpxor		$Zlo,$Xi,$Xi		# collect result
1280	  vpshufb	$bswap,$Ij,$Ij
1281	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1282	vxorps		$Zhi,$Xo,$Xo
1283	  vmovdqu	0x10-0x40($Htbl),$Hkey	# $Hkey^2
1284	 vpunpckhqdq	$Ij,$Ij,$T1
1285	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1286	vpxor		$Zmi,$Tred,$Tred
1287	 vxorps		$Ij,$T1,$T1
1288
1289	  vmovdqu	0x50($inp),$Ii		# I[5]
1290	vpxor		$Xi,$Tred,$Tred		# aggregated Karatsuba post-processing
1291	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1292	vpxor		$Xo,$Tred,$Tred
1293	vpslldq		\$8,$Tred,$T2
1294	 vpxor		$Xlo,$Zlo,$Zlo
1295	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1296	vpsrldq		\$8,$Tred,$Tred
1297	vpxor		$T2, $Xi, $Xi
1298	  vmovdqu	0x30-0x40($Htbl),$Hkey	# $Hkey^3
1299	  vpshufb	$bswap,$Ii,$Ii
1300	vxorps		$Tred,$Xo, $Xo
1301	 vpxor		$Xhi,$Zhi,$Zhi
1302	 vpunpckhqdq	$Ii,$Ii,$T2
1303	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1304	  vmovdqu	0x50-0x40($Htbl),$HK
1305	 vpxor		$Ii,$T2,$T2
1306	 vpxor		$Xmi,$Zmi,$Zmi
1307
1308	  vmovdqu	0x40($inp),$Ij		# I[4]
1309	vpalignr	\$8,$Xi,$Xi,$Tred	# 1st phase
1310	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1311	  vpshufb	$bswap,$Ij,$Ij
1312	 vpxor		$Zlo,$Xlo,$Xlo
1313	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1314	  vmovdqu	0x40-0x40($Htbl),$Hkey	# $Hkey^4
1315	 vpunpckhqdq	$Ij,$Ij,$T1
1316	 vpxor		$Zhi,$Xhi,$Xhi
1317	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1318	 vxorps		$Ij,$T1,$T1
1319	 vpxor		$Zmi,$Xmi,$Xmi
1320
1321	  vmovdqu	0x30($inp),$Ii		# I[3]
1322	vpclmulqdq	\$0x10,(%r10),$Xi,$Xi
1323	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1324	  vpshufb	$bswap,$Ii,$Ii
1325	 vpxor		$Xlo,$Zlo,$Zlo
1326	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1327	  vmovdqu	0x60-0x40($Htbl),$Hkey	# $Hkey^5
1328	 vpunpckhqdq	$Ii,$Ii,$T2
1329	 vpxor		$Xhi,$Zhi,$Zhi
1330	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1331	  vmovdqu	0x80-0x40($Htbl),$HK
1332	 vpxor		$Ii,$T2,$T2
1333	 vpxor		$Xmi,$Zmi,$Zmi
1334
1335	  vmovdqu	0x20($inp),$Ij		# I[2]
1336	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1337	  vpshufb	$bswap,$Ij,$Ij
1338	 vpxor		$Zlo,$Xlo,$Xlo
1339	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1340	  vmovdqu	0x70-0x40($Htbl),$Hkey	# $Hkey^6
1341	 vpunpckhqdq	$Ij,$Ij,$T1
1342	 vpxor		$Zhi,$Xhi,$Xhi
1343	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1344	 vpxor		$Ij,$T1,$T1
1345	 vpxor		$Zmi,$Xmi,$Xmi
1346	vxorps		$Tred,$Xi,$Xi
1347
1348	  vmovdqu	0x10($inp),$Ii		# I[1]
1349	vpalignr	\$8,$Xi,$Xi,$Tred	# 2nd phase
1350	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1351	  vpshufb	$bswap,$Ii,$Ii
1352	 vpxor		$Xlo,$Zlo,$Zlo
1353	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1354	  vmovdqu	0x90-0x40($Htbl),$Hkey	# $Hkey^7
1355	vpclmulqdq	\$0x10,(%r10),$Xi,$Xi
1356	vxorps		$Xo,$Tred,$Tred
1357	 vpunpckhqdq	$Ii,$Ii,$T2
1358	 vpxor		$Xhi,$Zhi,$Zhi
1359	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1360	  vmovdqu	0xb0-0x40($Htbl),$HK
1361	 vpxor		$Ii,$T2,$T2
1362	 vpxor		$Xmi,$Zmi,$Zmi
1363
1364	  vmovdqu	($inp),$Ij		# I[0]
1365	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1366	  vpshufb	$bswap,$Ij,$Ij
1367	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1368	  vmovdqu	0xa0-0x40($Htbl),$Hkey	# $Hkey^8
1369	vpxor		$Tred,$Ij,$Ij
1370	 vpclmulqdq	\$0x10,$HK,  $T2,$Xmi
1371	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1372
1373	lea		0x80($inp),$inp
1374	sub		\$0x80,$len
1375	jnc		.Loop8x_avx
1376
1377	add		\$0x80,$len
1378	jmp		.Ltail_no_xor_avx
1379
1380.align	32
1381.Lshort_avx:
1382	vmovdqu		-0x10($inp,$len),$Ii	# very last word
1383	lea		($inp,$len),$inp
1384	vmovdqu		0x00-0x40($Htbl),$Hkey	# $Hkey^1
1385	vmovdqu		0x20-0x40($Htbl),$HK
1386	vpshufb		$bswap,$Ii,$Ij
1387
1388	vmovdqa		$Xlo,$Zlo		# subtle way to zero $Zlo,
1389	vmovdqa		$Xhi,$Zhi		# $Zhi and
1390	vmovdqa		$Xmi,$Zmi		# $Zmi
1391	sub		\$0x10,$len
1392	jz		.Ltail_avx
1393
1394	vpunpckhqdq	$Ij,$Ij,$T1
1395	vpxor		$Xlo,$Zlo,$Zlo
1396	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1397	vpxor		$Ij,$T1,$T1
1398	 vmovdqu	-0x20($inp),$Ii
1399	vpxor		$Xhi,$Zhi,$Zhi
1400	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1401	vmovdqu		0x10-0x40($Htbl),$Hkey	# $Hkey^2
1402	 vpshufb	$bswap,$Ii,$Ij
1403	vpxor		$Xmi,$Zmi,$Zmi
1404	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1405	vpsrldq		\$8,$HK,$HK
1406	sub		\$0x10,$len
1407	jz		.Ltail_avx
1408
1409	vpunpckhqdq	$Ij,$Ij,$T1
1410	vpxor		$Xlo,$Zlo,$Zlo
1411	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1412	vpxor		$Ij,$T1,$T1
1413	 vmovdqu	-0x30($inp),$Ii
1414	vpxor		$Xhi,$Zhi,$Zhi
1415	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1416	vmovdqu		0x30-0x40($Htbl),$Hkey	# $Hkey^3
1417	 vpshufb	$bswap,$Ii,$Ij
1418	vpxor		$Xmi,$Zmi,$Zmi
1419	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1420	vmovdqu		0x50-0x40($Htbl),$HK
1421	sub		\$0x10,$len
1422	jz		.Ltail_avx
1423
1424	vpunpckhqdq	$Ij,$Ij,$T1
1425	vpxor		$Xlo,$Zlo,$Zlo
1426	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1427	vpxor		$Ij,$T1,$T1
1428	 vmovdqu	-0x40($inp),$Ii
1429	vpxor		$Xhi,$Zhi,$Zhi
1430	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1431	vmovdqu		0x40-0x40($Htbl),$Hkey	# $Hkey^4
1432	 vpshufb	$bswap,$Ii,$Ij
1433	vpxor		$Xmi,$Zmi,$Zmi
1434	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1435	vpsrldq		\$8,$HK,$HK
1436	sub		\$0x10,$len
1437	jz		.Ltail_avx
1438
1439	vpunpckhqdq	$Ij,$Ij,$T1
1440	vpxor		$Xlo,$Zlo,$Zlo
1441	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1442	vpxor		$Ij,$T1,$T1
1443	 vmovdqu	-0x50($inp),$Ii
1444	vpxor		$Xhi,$Zhi,$Zhi
1445	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1446	vmovdqu		0x60-0x40($Htbl),$Hkey	# $Hkey^5
1447	 vpshufb	$bswap,$Ii,$Ij
1448	vpxor		$Xmi,$Zmi,$Zmi
1449	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1450	vmovdqu		0x80-0x40($Htbl),$HK
1451	sub		\$0x10,$len
1452	jz		.Ltail_avx
1453
1454	vpunpckhqdq	$Ij,$Ij,$T1
1455	vpxor		$Xlo,$Zlo,$Zlo
1456	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1457	vpxor		$Ij,$T1,$T1
1458	 vmovdqu	-0x60($inp),$Ii
1459	vpxor		$Xhi,$Zhi,$Zhi
1460	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1461	vmovdqu		0x70-0x40($Htbl),$Hkey	# $Hkey^6
1462	 vpshufb	$bswap,$Ii,$Ij
1463	vpxor		$Xmi,$Zmi,$Zmi
1464	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1465	vpsrldq		\$8,$HK,$HK
1466	sub		\$0x10,$len
1467	jz		.Ltail_avx
1468
1469	vpunpckhqdq	$Ij,$Ij,$T1
1470	vpxor		$Xlo,$Zlo,$Zlo
1471	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1472	vpxor		$Ij,$T1,$T1
1473	 vmovdqu	-0x70($inp),$Ii
1474	vpxor		$Xhi,$Zhi,$Zhi
1475	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1476	vmovdqu		0x90-0x40($Htbl),$Hkey	# $Hkey^7
1477	 vpshufb	$bswap,$Ii,$Ij
1478	vpxor		$Xmi,$Zmi,$Zmi
1479	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1480	vmovq		0xb8-0x40($Htbl),$HK
1481	sub		\$0x10,$len
1482	jmp		.Ltail_avx
1483
1484.align	32
1485.Ltail_avx:
1486	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1487.Ltail_no_xor_avx:
1488	vpunpckhqdq	$Ij,$Ij,$T1
1489	vpxor		$Xlo,$Zlo,$Zlo
1490	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1491	vpxor		$Ij,$T1,$T1
1492	vpxor		$Xhi,$Zhi,$Zhi
1493	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1494	vpxor		$Xmi,$Zmi,$Zmi
1495	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1496
1497	vmovdqu		(%r10),$Tred
1498
1499	vpxor		$Xlo,$Zlo,$Xi
1500	vpxor		$Xhi,$Zhi,$Xo
1501	vpxor		$Xmi,$Zmi,$Zmi
1502
1503	vpxor		$Xi, $Zmi,$Zmi		# aggregated Karatsuba post-processing
1504	vpxor		$Xo, $Zmi,$Zmi
1505	vpslldq		\$8, $Zmi,$T2
1506	vpsrldq		\$8, $Zmi,$Zmi
1507	vpxor		$T2, $Xi, $Xi
1508	vpxor		$Zmi,$Xo, $Xo
1509
1510	vpclmulqdq	\$0x10,$Tred,$Xi,$T2	# 1st phase
1511	vpalignr	\$8,$Xi,$Xi,$Xi
1512	vpxor		$T2,$Xi,$Xi
1513
1514	vpclmulqdq	\$0x10,$Tred,$Xi,$T2	# 2nd phase
1515	vpalignr	\$8,$Xi,$Xi,$Xi
1516	vpxor		$Xo,$Xi,$Xi
1517	vpxor		$T2,$Xi,$Xi
1518
1519	cmp		\$0,$len
1520	jne		.Lshort_avx
1521
1522	vpshufb		$bswap,$Xi,$Xi
1523	vmovdqu		$Xi,($Xip)
1524	vzeroupper
1525___
1526$code.=<<___ if ($win64);
1527	movaps	(%rsp),%xmm6
1528	movaps	0x10(%rsp),%xmm7
1529	movaps	0x20(%rsp),%xmm8
1530	movaps	0x30(%rsp),%xmm9
1531	movaps	0x40(%rsp),%xmm10
1532	movaps	0x50(%rsp),%xmm11
1533	movaps	0x60(%rsp),%xmm12
1534	movaps	0x70(%rsp),%xmm13
1535	movaps	0x80(%rsp),%xmm14
1536	movaps	0x90(%rsp),%xmm15
1537	lea	0xa8(%rsp),%rsp
1538.LSEH_end_gcm_ghash_avx:
1539___
1540$code.=<<___;
1541	ret
1542.size	gcm_ghash_avx,.-gcm_ghash_avx
1543___
1544} else {
1545$code.=<<___;
1546	jmp	.L_ghash_clmul
1547.size	gcm_ghash_avx,.-gcm_ghash_avx
1548___
1549}
1550
1551$code.=<<___;
1552.align	64
1553.Lbswap_mask:
1554	.byte	15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1555.L0x1c2_polynomial:
1556	.byte	1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1557.L7_mask:
1558	.long	7,0,7,0
1559.L7_mask_poly:
1560	.long	7,0,`0xE1<<1`,0
1561.align	64
1562.type	.Lrem_4bit,\@object
1563.Lrem_4bit:
1564	.long	0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1565	.long	0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1566	.long	0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1567	.long	0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1568.type	.Lrem_8bit,\@object
1569.Lrem_8bit:
1570	.value	0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1571	.value	0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1572	.value	0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1573	.value	0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1574	.value	0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1575	.value	0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1576	.value	0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1577	.value	0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1578	.value	0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1579	.value	0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1580	.value	0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1581	.value	0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1582	.value	0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1583	.value	0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1584	.value	0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1585	.value	0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1586	.value	0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1587	.value	0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1588	.value	0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1589	.value	0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1590	.value	0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1591	.value	0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1592	.value	0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1593	.value	0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1594	.value	0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1595	.value	0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1596	.value	0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1597	.value	0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1598	.value	0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1599	.value	0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1600	.value	0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1601	.value	0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1602
1603.asciz	"GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1604.align	64
1605___
1606
1607# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1608#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
1609if ($win64) {
1610$rec="%rcx";
1611$frame="%rdx";
1612$context="%r8";
1613$disp="%r9";
1614
1615$code.=<<___;
1616.extern	__imp_RtlVirtualUnwind
1617.type	se_handler,\@abi-omnipotent
1618.align	16
1619se_handler:
1620	push	%rsi
1621	push	%rdi
1622	push	%rbx
1623	push	%rbp
1624	push	%r12
1625	push	%r13
1626	push	%r14
1627	push	%r15
1628	pushfq
1629	sub	\$64,%rsp
1630
1631	mov	120($context),%rax	# pull context->Rax
1632	mov	248($context),%rbx	# pull context->Rip
1633
1634	mov	8($disp),%rsi		# disp->ImageBase
1635	mov	56($disp),%r11		# disp->HandlerData
1636
1637	mov	0(%r11),%r10d		# HandlerData[0]
1638	lea	(%rsi,%r10),%r10	# prologue label
1639	cmp	%r10,%rbx		# context->Rip<prologue label
1640	jb	.Lin_prologue
1641
1642	mov	152($context),%rax	# pull context->Rsp
1643
1644	mov	4(%r11),%r10d		# HandlerData[1]
1645	lea	(%rsi,%r10),%r10	# epilogue label
1646	cmp	%r10,%rbx		# context->Rip>=epilogue label
1647	jae	.Lin_prologue
1648
1649	lea	48+280(%rax),%rax	# adjust "rsp"
1650
1651	mov	-8(%rax),%rbx
1652	mov	-16(%rax),%rbp
1653	mov	-24(%rax),%r12
1654	mov	-32(%rax),%r13
1655	mov	-40(%rax),%r14
1656	mov	-48(%rax),%r15
1657	mov	%rbx,144($context)	# restore context->Rbx
1658	mov	%rbp,160($context)	# restore context->Rbp
1659	mov	%r12,216($context)	# restore context->R12
1660	mov	%r13,224($context)	# restore context->R13
1661	mov	%r14,232($context)	# restore context->R14
1662	mov	%r15,240($context)	# restore context->R15
1663
1664.Lin_prologue:
1665	mov	8(%rax),%rdi
1666	mov	16(%rax),%rsi
1667	mov	%rax,152($context)	# restore context->Rsp
1668	mov	%rsi,168($context)	# restore context->Rsi
1669	mov	%rdi,176($context)	# restore context->Rdi
1670
1671	mov	40($disp),%rdi		# disp->ContextRecord
1672	mov	$context,%rsi		# context
1673	mov	\$`1232/8`,%ecx		# sizeof(CONTEXT)
1674	.long	0xa548f3fc		# cld; rep movsq
1675
1676	mov	$disp,%rsi
1677	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
1678	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
1679	mov	0(%rsi),%r8		# arg3, disp->ControlPc
1680	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
1681	mov	40(%rsi),%r10		# disp->ContextRecord
1682	lea	56(%rsi),%r11		# &disp->HandlerData
1683	lea	24(%rsi),%r12		# &disp->EstablisherFrame
1684	mov	%r10,32(%rsp)		# arg5
1685	mov	%r11,40(%rsp)		# arg6
1686	mov	%r12,48(%rsp)		# arg7
1687	mov	%rcx,56(%rsp)		# arg8, (NULL)
1688	call	*__imp_RtlVirtualUnwind(%rip)
1689
1690	mov	\$1,%eax		# ExceptionContinueSearch
1691	add	\$64,%rsp
1692	popfq
1693	pop	%r15
1694	pop	%r14
1695	pop	%r13
1696	pop	%r12
1697	pop	%rbp
1698	pop	%rbx
1699	pop	%rdi
1700	pop	%rsi
1701	ret
1702.size	se_handler,.-se_handler
1703
1704.section	.pdata
1705.align	4
1706	.rva	.LSEH_begin_gcm_gmult_4bit
1707	.rva	.LSEH_end_gcm_gmult_4bit
1708	.rva	.LSEH_info_gcm_gmult_4bit
1709
1710	.rva	.LSEH_begin_gcm_ghash_4bit
1711	.rva	.LSEH_end_gcm_ghash_4bit
1712	.rva	.LSEH_info_gcm_ghash_4bit
1713
1714	.rva	.LSEH_begin_gcm_init_clmul
1715	.rva	.LSEH_end_gcm_init_clmul
1716	.rva	.LSEH_info_gcm_init_clmul
1717
1718	.rva	.LSEH_begin_gcm_ghash_clmul
1719	.rva	.LSEH_end_gcm_ghash_clmul
1720	.rva	.LSEH_info_gcm_ghash_clmul
1721___
1722$code.=<<___	if ($avx);
1723	.rva	.LSEH_begin_gcm_init_avx
1724	.rva	.LSEH_end_gcm_init_avx
1725	.rva	.LSEH_info_gcm_init_clmul
1726
1727	.rva	.LSEH_begin_gcm_ghash_avx
1728	.rva	.LSEH_end_gcm_ghash_avx
1729	.rva	.LSEH_info_gcm_ghash_clmul
1730___
1731$code.=<<___;
1732.section	.xdata
1733.align	8
1734.LSEH_info_gcm_gmult_4bit:
1735	.byte	9,0,0,0
1736	.rva	se_handler
1737	.rva	.Lgmult_prologue,.Lgmult_epilogue	# HandlerData
1738.LSEH_info_gcm_ghash_4bit:
1739	.byte	9,0,0,0
1740	.rva	se_handler
1741	.rva	.Lghash_prologue,.Lghash_epilogue	# HandlerData
1742.LSEH_info_gcm_init_clmul:
1743	.byte	0x01,0x08,0x03,0x00
1744	.byte	0x08,0x68,0x00,0x00	#movaps	0x00(rsp),xmm6
1745	.byte	0x04,0x22,0x00,0x00	#sub	rsp,0x18
1746.LSEH_info_gcm_ghash_clmul:
1747	.byte	0x01,0x33,0x16,0x00
1748	.byte	0x33,0xf8,0x09,0x00	#movaps 0x90(rsp),xmm15
1749	.byte	0x2e,0xe8,0x08,0x00	#movaps 0x80(rsp),xmm14
1750	.byte	0x29,0xd8,0x07,0x00	#movaps 0x70(rsp),xmm13
1751	.byte	0x24,0xc8,0x06,0x00	#movaps 0x60(rsp),xmm12
1752	.byte	0x1f,0xb8,0x05,0x00	#movaps 0x50(rsp),xmm11
1753	.byte	0x1a,0xa8,0x04,0x00	#movaps 0x40(rsp),xmm10
1754	.byte	0x15,0x98,0x03,0x00	#movaps 0x30(rsp),xmm9
1755	.byte	0x10,0x88,0x02,0x00	#movaps 0x20(rsp),xmm8
1756	.byte	0x0c,0x78,0x01,0x00	#movaps 0x10(rsp),xmm7
1757	.byte	0x08,0x68,0x00,0x00	#movaps 0x00(rsp),xmm6
1758	.byte	0x04,0x01,0x15,0x00	#sub	rsp,0xa8
1759___
1760}
1761
1762$code =~ s/\`([^\`]*)\`/eval($1)/gem;
1763
1764print $code;
1765
1766close STDOUT;
1767