xref: /openssl/crypto/modes/asm/ghash-x86_64.pl (revision cd84d883) 
1#! /usr/bin/env perl
2# Copyright 2010-2020 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. 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 for providing access to a
48# Westmere-based system on behalf of Intel Open Source Technology Centre.
49
50# December 2012
51#
52# Overhaul: aggregate Karatsuba post-processing, improve ILP in
53# reduction_alg9, increase reduction aggregate factor to 4x. As for
54# the latter. ghash-x86.pl discusses that it makes lesser sense to
55# increase aggregate factor. Then why increase here? Critical path
56# consists of 3 independent pclmulqdq instructions, Karatsuba post-
57# processing and reduction. "On top" of this we lay down aggregated
58# multiplication operations, triplets of independent pclmulqdq's. As
59# issue rate for pclmulqdq is limited, it makes lesser sense to
60# aggregate more multiplications than it takes to perform remaining
61# non-multiplication operations. 2x is near-optimal coefficient for
62# contemporary Intel CPUs (therefore modest improvement coefficient),
63# but not for Bulldozer. Latter is because logical SIMD operations
64# are twice as slow in comparison to Intel, so that critical path is
65# longer. A CPU with higher pclmulqdq issue rate would also benefit
66# from higher aggregate factor...
67#
68# Westmere	1.78(+13%)
69# Sandy Bridge	1.80(+8%)
70# Ivy Bridge	1.80(+7%)
71# Haswell	0.55(+93%) (if system doesn't support AVX)
72# Broadwell	0.45(+110%)(if system doesn't support AVX)
73# Skylake	0.44(+110%)(if system doesn't support AVX)
74# Bulldozer	1.49(+27%)
75# Silvermont	2.88(+13%)
76# Knights L	2.12(-)    (if system doesn't support AVX)
77# Goldmont	1.08(+24%)
78
79# March 2013
80#
81# ... 8x aggregate factor AVX code path is using reduction algorithm
82# suggested by Shay Gueron[1]. Even though contemporary AVX-capable
83# CPUs such as Sandy and Ivy Bridge can execute it, the code performs
84# sub-optimally in comparison to above mentioned version. But thanks
85# to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
86# it performs in 0.41 cycles per byte on Haswell processor, in
87# 0.29 on Broadwell, and in 0.36 on Skylake.
88#
89# Knights Landing achieves 1.09 cpb.
90#
91# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
92
93# $output is the last argument if it looks like a file (it has an extension)
94# $flavour is the first argument if it doesn't look like a file
95$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
96$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
97
98$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 $xlate) or
103die "can't locate x86_64-xlate.pl";
104
105if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
106		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
107	$avx = ($1>=2.20) + ($1>=2.22);
108}
109
110if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
111	    `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
112	$avx = ($1>=2.09) + ($1>=2.10);
113}
114
115if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
116	    `ml64 2>&1` =~ /Version ([0-9]+)\./) {
117	$avx = ($1>=10) + ($1>=11);
118}
119
120if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:clang|LLVM) version|.*based on LLVM) ([0-9]+\.[0-9]+)/) {
121	$avx = ($2>=3.0) + ($2>3.0);
122}
123
124open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
125    or die "can't call $xlate: $!";
126*STDOUT=*OUT;
127
128$do4xaggr=1;
129
130# common register layout
131$nlo="%rax";
132$nhi="%rbx";
133$Zlo="%r8";
134$Zhi="%r9";
135$tmp="%r10";
136$rem_4bit = "%r11";
137
138$Xi="%rdi";
139$Htbl="%rsi";
140
141# per-function register layout
142$cnt="%rcx";
143$rem="%rdx";
144
145sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/	or
146			$r =~ s/%[er]([sd]i)/%\1l/	or
147			$r =~ s/%[er](bp)/%\1l/		or
148			$r =~ s/%(r[0-9]+)[d]?/%\1b/;   $r; }
149
150sub AUTOLOAD()		# thunk [simplified] 32-bit style perlasm
151{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
152  my $arg = pop;
153    $arg = "\$$arg" if ($arg*1 eq $arg);
154    $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
155}
156
157{ my $N;
158  sub loop() {
159  my $inp = shift;
160
161	$N++;
162$code.=<<___;
163	xor	$nlo,$nlo
164	xor	$nhi,$nhi
165	mov	`&LB("$Zlo")`,`&LB("$nlo")`
166	mov	`&LB("$Zlo")`,`&LB("$nhi")`
167	shl	\$4,`&LB("$nlo")`
168	mov	\$14,$cnt
169	mov	8($Htbl,$nlo),$Zlo
170	mov	($Htbl,$nlo),$Zhi
171	and	\$0xf0,`&LB("$nhi")`
172	mov	$Zlo,$rem
173	jmp	.Loop$N
174
175.align	16
176.Loop$N:
177	shr	\$4,$Zlo
178	and	\$0xf,$rem
179	mov	$Zhi,$tmp
180	mov	($inp,$cnt),`&LB("$nlo")`
181	shr	\$4,$Zhi
182	xor	8($Htbl,$nhi),$Zlo
183	shl	\$60,$tmp
184	xor	($Htbl,$nhi),$Zhi
185	mov	`&LB("$nlo")`,`&LB("$nhi")`
186	xor	($rem_4bit,$rem,8),$Zhi
187	mov	$Zlo,$rem
188	shl	\$4,`&LB("$nlo")`
189	xor	$tmp,$Zlo
190	dec	$cnt
191	js	.Lbreak$N
192
193	shr	\$4,$Zlo
194	and	\$0xf,$rem
195	mov	$Zhi,$tmp
196	shr	\$4,$Zhi
197	xor	8($Htbl,$nlo),$Zlo
198	shl	\$60,$tmp
199	xor	($Htbl,$nlo),$Zhi
200	and	\$0xf0,`&LB("$nhi")`
201	xor	($rem_4bit,$rem,8),$Zhi
202	mov	$Zlo,$rem
203	xor	$tmp,$Zlo
204	jmp	.Loop$N
205
206.align	16
207.Lbreak$N:
208	shr	\$4,$Zlo
209	and	\$0xf,$rem
210	mov	$Zhi,$tmp
211	shr	\$4,$Zhi
212	xor	8($Htbl,$nlo),$Zlo
213	shl	\$60,$tmp
214	xor	($Htbl,$nlo),$Zhi
215	and	\$0xf0,`&LB("$nhi")`
216	xor	($rem_4bit,$rem,8),$Zhi
217	mov	$Zlo,$rem
218	xor	$tmp,$Zlo
219
220	shr	\$4,$Zlo
221	and	\$0xf,$rem
222	mov	$Zhi,$tmp
223	shr	\$4,$Zhi
224	xor	8($Htbl,$nhi),$Zlo
225	shl	\$60,$tmp
226	xor	($Htbl,$nhi),$Zhi
227	xor	$tmp,$Zlo
228	xor	($rem_4bit,$rem,8),$Zhi
229
230	bswap	$Zlo
231	bswap	$Zhi
232___
233}}
234
235$code=<<___;
236.text
237.extern	OPENSSL_ia32cap_P
238
239.globl	gcm_gmult_4bit
240.type	gcm_gmult_4bit,\@function,2
241.align	16
242gcm_gmult_4bit:
243.cfi_startproc
244	endbranch
245	push	%rbx
246.cfi_push	%rbx
247	push	%rbp		# %rbp and others are pushed exclusively in
248.cfi_push	%rbp
249	push	%r12		# order to reuse Win64 exception handler...
250.cfi_push	%r12
251	push	%r13
252.cfi_push	%r13
253	push	%r14
254.cfi_push	%r14
255	push	%r15
256.cfi_push	%r15
257	sub	\$280,%rsp
258.cfi_adjust_cfa_offset	280
259.Lgmult_prologue:
260
261	movzb	15($Xi),$Zlo
262	lea	.Lrem_4bit(%rip),$rem_4bit
263___
264	&loop	($Xi);
265$code.=<<___;
266	mov	$Zlo,8($Xi)
267	mov	$Zhi,($Xi)
268
269	lea	280+48(%rsp),%rsi
270.cfi_def_cfa	%rsi,8
271	mov	-8(%rsi),%rbx
272.cfi_restore	%rbx
273	lea	(%rsi),%rsp
274.cfi_def_cfa_register	%rsp
275.Lgmult_epilogue:
276	ret
277.cfi_endproc
278.size	gcm_gmult_4bit,.-gcm_gmult_4bit
279___
280
281# per-function register layout
282$inp="%rdx";
283$len="%rcx";
284$rem_8bit=$rem_4bit;
285
286$code.=<<___;
287.globl	gcm_ghash_4bit
288.type	gcm_ghash_4bit,\@function,4
289.align	16
290gcm_ghash_4bit:
291.cfi_startproc
292	endbranch
293	push	%rbx
294.cfi_push	%rbx
295	push	%rbp
296.cfi_push	%rbp
297	push	%r12
298.cfi_push	%r12
299	push	%r13
300.cfi_push	%r13
301	push	%r14
302.cfi_push	%r14
303	push	%r15
304.cfi_push	%r15
305	sub	\$280,%rsp
306.cfi_adjust_cfa_offset	280
307.Lghash_prologue:
308	mov	$inp,%r14		# reassign couple of args
309	mov	$len,%r15
310___
311{ my $inp="%r14";
312  my $dat="%edx";
313  my $len="%r15";
314  my @nhi=("%ebx","%ecx");
315  my @rem=("%r12","%r13");
316  my $Hshr4="%rbp";
317
318	&sub	($Htbl,-128);		# size optimization
319	&lea	($Hshr4,"16+128(%rsp)");
320	{ my @lo =($nlo,$nhi);
321          my @hi =($Zlo,$Zhi);
322
323	  &xor	($dat,$dat);
324	  for ($i=0,$j=-2;$i<18;$i++,$j++) {
325	    &mov	("$j(%rsp)",&LB($dat))		if ($i>1);
326	    &or		($lo[0],$tmp)			if ($i>1);
327	    &mov	(&LB($dat),&LB($lo[1]))		if ($i>0 && $i<17);
328	    &shr	($lo[1],4)			if ($i>0 && $i<17);
329	    &mov	($tmp,$hi[1])			if ($i>0 && $i<17);
330	    &shr	($hi[1],4)			if ($i>0 && $i<17);
331	    &mov	("8*$j($Hshr4)",$hi[0])		if ($i>1);
332	    &mov	($hi[0],"16*$i+0-128($Htbl)")	if ($i<16);
333	    &shl	(&LB($dat),4)			if ($i>0 && $i<17);
334	    &mov	("8*$j-128($Hshr4)",$lo[0])	if ($i>1);
335	    &mov	($lo[0],"16*$i+8-128($Htbl)")	if ($i<16);
336	    &shl	($tmp,60)			if ($i>0 && $i<17);
337
338	    push	(@lo,shift(@lo));
339	    push	(@hi,shift(@hi));
340	  }
341	}
342	&add	($Htbl,-128);
343	&mov	($Zlo,"8($Xi)");
344	&mov	($Zhi,"0($Xi)");
345	&add	($len,$inp);		# pointer to the end of data
346	&lea	($rem_8bit,".Lrem_8bit(%rip)");
347	&jmp	(".Louter_loop");
348
349$code.=".align	16\n.Louter_loop:\n";
350	&xor	($Zhi,"($inp)");
351	&mov	("%rdx","8($inp)");
352	&lea	($inp,"16($inp)");
353	&xor	("%rdx",$Zlo);
354	&mov	("($Xi)",$Zhi);
355	&mov	("8($Xi)","%rdx");
356	&shr	("%rdx",32);
357
358	&xor	($nlo,$nlo);
359	&rol	($dat,8);
360	&mov	(&LB($nlo),&LB($dat));
361	&movz	($nhi[0],&LB($dat));
362	&shl	(&LB($nlo),4);
363	&shr	($nhi[0],4);
364
365	for ($j=11,$i=0;$i<15;$i++) {
366	    &rol	($dat,8);
367	    &xor	($Zlo,"8($Htbl,$nlo)")			if ($i>0);
368	    &xor	($Zhi,"($Htbl,$nlo)")			if ($i>0);
369	    &mov	($Zlo,"8($Htbl,$nlo)")			if ($i==0);
370	    &mov	($Zhi,"($Htbl,$nlo)")			if ($i==0);
371
372	    &mov	(&LB($nlo),&LB($dat));
373	    &xor	($Zlo,$tmp)				if ($i>0);
374	    &movzw	($rem[1],"($rem_8bit,$rem[1],2)")	if ($i>0);
375
376	    &movz	($nhi[1],&LB($dat));
377	    &shl	(&LB($nlo),4);
378	    &movzb	($rem[0],"(%rsp,$nhi[0])");
379
380	    &shr	($nhi[1],4)				if ($i<14);
381	    &and	($nhi[1],0xf0)				if ($i==14);
382	    &shl	($rem[1],48)				if ($i>0);
383	    &xor	($rem[0],$Zlo);
384
385	    &mov	($tmp,$Zhi);
386	    &xor	($Zhi,$rem[1])				if ($i>0);
387	    &shr	($Zlo,8);
388
389	    &movz	($rem[0],&LB($rem[0]));
390	    &mov	($dat,"$j($Xi)")			if (--$j%4==0);
391	    &shr	($Zhi,8);
392
393	    &xor	($Zlo,"-128($Hshr4,$nhi[0],8)");
394	    &shl	($tmp,56);
395	    &xor	($Zhi,"($Hshr4,$nhi[0],8)");
396
397	    unshift	(@nhi,pop(@nhi));		# "rotate" registers
398	    unshift	(@rem,pop(@rem));
399	}
400	&movzw	($rem[1],"($rem_8bit,$rem[1],2)");
401	&xor	($Zlo,"8($Htbl,$nlo)");
402	&xor	($Zhi,"($Htbl,$nlo)");
403
404	&shl	($rem[1],48);
405	&xor	($Zlo,$tmp);
406
407	&xor	($Zhi,$rem[1]);
408	&movz	($rem[0],&LB($Zlo));
409	&shr	($Zlo,4);
410
411	&mov	($tmp,$Zhi);
412	&shl	(&LB($rem[0]),4);
413	&shr	($Zhi,4);
414
415	&xor	($Zlo,"8($Htbl,$nhi[0])");
416	&movzw	($rem[0],"($rem_8bit,$rem[0],2)");
417	&shl	($tmp,60);
418
419	&xor	($Zhi,"($Htbl,$nhi[0])");
420	&xor	($Zlo,$tmp);
421	&shl	($rem[0],48);
422
423	&bswap	($Zlo);
424	&xor	($Zhi,$rem[0]);
425
426	&bswap	($Zhi);
427	&cmp	($inp,$len);
428	&jb	(".Louter_loop");
429}
430$code.=<<___;
431	mov	$Zlo,8($Xi)
432	mov	$Zhi,($Xi)
433
434	lea	280+48(%rsp),%rsi
435.cfi_def_cfa	%rsi,8
436	mov	-48(%rsi),%r15
437.cfi_restore	%r15
438	mov	-40(%rsi),%r14
439.cfi_restore	%r14
440	mov	-32(%rsi),%r13
441.cfi_restore	%r13
442	mov	-24(%rsi),%r12
443.cfi_restore	%r12
444	mov	-16(%rsi),%rbp
445.cfi_restore	%rbp
446	mov	-8(%rsi),%rbx
447.cfi_restore	%rbx
448	lea	0(%rsi),%rsp
449.cfi_def_cfa_register	%rsp
450.Lghash_epilogue:
451	ret
452.cfi_endproc
453.size	gcm_ghash_4bit,.-gcm_ghash_4bit
454___
455
456######################################################################
457# PCLMULQDQ version.
458
459@_4args=$win64?	("%rcx","%rdx","%r8", "%r9") :	# Win64 order
460		("%rdi","%rsi","%rdx","%rcx");	# Unix order
461
462($Xi,$Xhi)=("%xmm0","%xmm1");	$Hkey="%xmm2";
463($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
464
465sub clmul64x64_T2 {	# minimal register pressure
466my ($Xhi,$Xi,$Hkey,$HK)=@_;
467
468if (!defined($HK)) {	$HK = $T2;
469$code.=<<___;
470	movdqa		$Xi,$Xhi		#
471	pshufd		\$0b01001110,$Xi,$T1
472	pshufd		\$0b01001110,$Hkey,$T2
473	pxor		$Xi,$T1			#
474	pxor		$Hkey,$T2
475___
476} else {
477$code.=<<___;
478	movdqa		$Xi,$Xhi		#
479	pshufd		\$0b01001110,$Xi,$T1
480	pxor		$Xi,$T1			#
481___
482}
483$code.=<<___;
484	pclmulqdq	\$0x00,$Hkey,$Xi	#######
485	pclmulqdq	\$0x11,$Hkey,$Xhi	#######
486	pclmulqdq	\$0x00,$HK,$T1		#######
487	pxor		$Xi,$T1			#
488	pxor		$Xhi,$T1		#
489
490	movdqa		$T1,$T2			#
491	psrldq		\$8,$T1
492	pslldq		\$8,$T2			#
493	pxor		$T1,$Xhi
494	pxor		$T2,$Xi			#
495___
496}
497
498sub reduction_alg9 {	# 17/11 times faster than Intel version
499my ($Xhi,$Xi) = @_;
500
501$code.=<<___;
502	# 1st phase
503	movdqa		$Xi,$T2			#
504	movdqa		$Xi,$T1
505	psllq		\$5,$Xi
506	pxor		$Xi,$T1			#
507	psllq		\$1,$Xi
508	pxor		$T1,$Xi			#
509	psllq		\$57,$Xi		#
510	movdqa		$Xi,$T1			#
511	pslldq		\$8,$Xi
512	psrldq		\$8,$T1			#
513	pxor		$T2,$Xi
514	pxor		$T1,$Xhi		#
515
516	# 2nd phase
517	movdqa		$Xi,$T2
518	psrlq		\$1,$Xi
519	pxor		$T2,$Xhi		#
520	pxor		$Xi,$T2
521	psrlq		\$5,$Xi
522	pxor		$T2,$Xi			#
523	psrlq		\$1,$Xi			#
524	pxor		$Xhi,$Xi		#
525___
526}
527
528{ my ($Htbl,$Xip)=@_4args;
529  my $HK="%xmm6";
530
531$code.=<<___;
532.globl	gcm_init_clmul
533.type	gcm_init_clmul,\@abi-omnipotent
534.align	16
535gcm_init_clmul:
536.cfi_startproc
537.L_init_clmul:
538___
539$code.=<<___ if ($win64);
540.LSEH_begin_gcm_init_clmul:
541	# I can't trust assembler to use specific encoding:-(
542	.byte	0x48,0x83,0xec,0x18		#sub	$0x18,%rsp
543	.byte	0x0f,0x29,0x34,0x24		#movaps	%xmm6,(%rsp)
544___
545$code.=<<___;
546	movdqu		($Xip),$Hkey
547	pshufd		\$0b01001110,$Hkey,$Hkey	# dword swap
548
549	# <<1 twist
550	pshufd		\$0b11111111,$Hkey,$T2	# broadcast uppermost dword
551	movdqa		$Hkey,$T1
552	psllq		\$1,$Hkey
553	pxor		$T3,$T3			#
554	psrlq		\$63,$T1
555	pcmpgtd		$T2,$T3			# broadcast carry bit
556	pslldq		\$8,$T1
557	por		$T1,$Hkey		# H<<=1
558
559	# magic reduction
560	pand		.L0x1c2_polynomial(%rip),$T3
561	pxor		$T3,$Hkey		# if(carry) H^=0x1c2_polynomial
562
563	# calculate H^2
564	pshufd		\$0b01001110,$Hkey,$HK
565	movdqa		$Hkey,$Xi
566	pxor		$Hkey,$HK
567___
568	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);
569	&reduction_alg9	($Xhi,$Xi);
570$code.=<<___;
571	pshufd		\$0b01001110,$Hkey,$T1
572	pshufd		\$0b01001110,$Xi,$T2
573	pxor		$Hkey,$T1		# Karatsuba pre-processing
574	movdqu		$Hkey,0x00($Htbl)	# save H
575	pxor		$Xi,$T2			# Karatsuba pre-processing
576	movdqu		$Xi,0x10($Htbl)		# save H^2
577	palignr		\$8,$T1,$T2		# low part is H.lo^H.hi...
578	movdqu		$T2,0x20($Htbl)		# save Karatsuba "salt"
579___
580if ($do4xaggr) {
581	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H^3
582	&reduction_alg9	($Xhi,$Xi);
583$code.=<<___;
584	movdqa		$Xi,$T3
585___
586	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H^4
587	&reduction_alg9	($Xhi,$Xi);
588$code.=<<___;
589	pshufd		\$0b01001110,$T3,$T1
590	pshufd		\$0b01001110,$Xi,$T2
591	pxor		$T3,$T1			# Karatsuba pre-processing
592	movdqu		$T3,0x30($Htbl)		# save H^3
593	pxor		$Xi,$T2			# Karatsuba pre-processing
594	movdqu		$Xi,0x40($Htbl)		# save H^4
595	palignr		\$8,$T1,$T2		# low part is H^3.lo^H^3.hi...
596	movdqu		$T2,0x50($Htbl)		# save Karatsuba "salt"
597___
598}
599$code.=<<___ if ($win64);
600	movaps	(%rsp),%xmm6
601	lea	0x18(%rsp),%rsp
602.LSEH_end_gcm_init_clmul:
603___
604$code.=<<___;
605	ret
606.cfi_endproc
607.size	gcm_init_clmul,.-gcm_init_clmul
608___
609}
610
611{ my ($Xip,$Htbl)=@_4args;
612
613$code.=<<___;
614.globl	gcm_gmult_clmul
615.type	gcm_gmult_clmul,\@abi-omnipotent
616.align	16
617gcm_gmult_clmul:
618.cfi_startproc
619	endbranch
620.L_gmult_clmul:
621	movdqu		($Xip),$Xi
622	movdqa		.Lbswap_mask(%rip),$T3
623	movdqu		($Htbl),$Hkey
624	movdqu		0x20($Htbl),$T2
625	pshufb		$T3,$Xi
626___
627	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$T2);
628$code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
629	# experimental alternative. special thing about is that there
630	# no dependency between the two multiplications...
631	mov		\$`0xE1<<1`,%eax
632	mov		\$0xA040608020C0E000,%r10	# ((7..0)·0xE0)&0xff
633	mov		\$0x07,%r11d
634	movq		%rax,$T1
635	movq		%r10,$T2
636	movq		%r11,$T3		# borrow $T3
637	pand		$Xi,$T3
638	pshufb		$T3,$T2			# ($Xi&7)·0xE0
639	movq		%rax,$T3
640	pclmulqdq	\$0x00,$Xi,$T1		# ·(0xE1<<1)
641	pxor		$Xi,$T2
642	pslldq		\$15,$T2
643	paddd		$T2,$T2			# <<(64+56+1)
644	pxor		$T2,$Xi
645	pclmulqdq	\$0x01,$T3,$Xi
646	movdqa		.Lbswap_mask(%rip),$T3	# reload $T3
647	psrldq		\$1,$T1
648	pxor		$T1,$Xhi
649	pslldq		\$7,$Xi
650	pxor		$Xhi,$Xi
651___
652$code.=<<___;
653	pshufb		$T3,$Xi
654	movdqu		$Xi,($Xip)
655	ret
656.cfi_endproc
657.size	gcm_gmult_clmul,.-gcm_gmult_clmul
658___
659}
660
661{ my ($Xip,$Htbl,$inp,$len)=@_4args;
662  my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
663  my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
664
665$code.=<<___;
666.globl	gcm_ghash_clmul
667.type	gcm_ghash_clmul,\@abi-omnipotent
668.align	32
669gcm_ghash_clmul:
670.cfi_startproc
671	endbranch
672.L_ghash_clmul:
673___
674$code.=<<___ if ($win64);
675	lea	-0x88(%rsp),%rax
676.LSEH_begin_gcm_ghash_clmul:
677	# I can't trust assembler to use specific encoding:-(
678	.byte	0x48,0x8d,0x60,0xe0		#lea	-0x20(%rax),%rsp
679	.byte	0x0f,0x29,0x70,0xe0		#movaps	%xmm6,-0x20(%rax)
680	.byte	0x0f,0x29,0x78,0xf0		#movaps	%xmm7,-0x10(%rax)
681	.byte	0x44,0x0f,0x29,0x00		#movaps	%xmm8,0(%rax)
682	.byte	0x44,0x0f,0x29,0x48,0x10	#movaps	%xmm9,0x10(%rax)
683	.byte	0x44,0x0f,0x29,0x50,0x20	#movaps	%xmm10,0x20(%rax)
684	.byte	0x44,0x0f,0x29,0x58,0x30	#movaps	%xmm11,0x30(%rax)
685	.byte	0x44,0x0f,0x29,0x60,0x40	#movaps	%xmm12,0x40(%rax)
686	.byte	0x44,0x0f,0x29,0x68,0x50	#movaps	%xmm13,0x50(%rax)
687	.byte	0x44,0x0f,0x29,0x70,0x60	#movaps	%xmm14,0x60(%rax)
688	.byte	0x44,0x0f,0x29,0x78,0x70	#movaps	%xmm15,0x70(%rax)
689___
690$code.=<<___;
691	movdqa		.Lbswap_mask(%rip),$T3
692
693	movdqu		($Xip),$Xi
694	movdqu		($Htbl),$Hkey
695	movdqu		0x20($Htbl),$HK
696	pshufb		$T3,$Xi
697
698	sub		\$0x10,$len
699	jz		.Lodd_tail
700
701	movdqu		0x10($Htbl),$Hkey2
702___
703if ($do4xaggr) {
704my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
705
706$code.=<<___;
707	mov		OPENSSL_ia32cap_P+4(%rip),%eax
708	cmp		\$0x30,$len
709	jb		.Lskip4x
710
711	and		\$`1<<26|1<<22`,%eax	# isolate MOVBE+XSAVE
712	cmp		\$`1<<22`,%eax		# check for MOVBE without XSAVE
713	je		.Lskip4x
714
715	sub		\$0x30,$len
716	mov		\$0xA040608020C0E000,%rax	# ((7..0)·0xE0)&0xff
717	movdqu		0x30($Htbl),$Hkey3
718	movdqu		0x40($Htbl),$Hkey4
719
720	#######
721	# Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
722	#
723	movdqu		0x30($inp),$Xln
724	 movdqu		0x20($inp),$Xl
725	pshufb		$T3,$Xln
726	 pshufb		$T3,$Xl
727	movdqa		$Xln,$Xhn
728	pshufd		\$0b01001110,$Xln,$Xmn
729	pxor		$Xln,$Xmn
730	pclmulqdq	\$0x00,$Hkey,$Xln
731	pclmulqdq	\$0x11,$Hkey,$Xhn
732	pclmulqdq	\$0x00,$HK,$Xmn
733
734	movdqa		$Xl,$Xh
735	pshufd		\$0b01001110,$Xl,$Xm
736	pxor		$Xl,$Xm
737	pclmulqdq	\$0x00,$Hkey2,$Xl
738	pclmulqdq	\$0x11,$Hkey2,$Xh
739	pclmulqdq	\$0x10,$HK,$Xm
740	xorps		$Xl,$Xln
741	xorps		$Xh,$Xhn
742	movups		0x50($Htbl),$HK
743	xorps		$Xm,$Xmn
744
745	movdqu		0x10($inp),$Xl
746	 movdqu		0($inp),$T1
747	pshufb		$T3,$Xl
748	 pshufb		$T3,$T1
749	movdqa		$Xl,$Xh
750	pshufd		\$0b01001110,$Xl,$Xm
751	 pxor		$T1,$Xi
752	pxor		$Xl,$Xm
753	pclmulqdq	\$0x00,$Hkey3,$Xl
754	 movdqa		$Xi,$Xhi
755	 pshufd		\$0b01001110,$Xi,$T1
756	 pxor		$Xi,$T1
757	pclmulqdq	\$0x11,$Hkey3,$Xh
758	pclmulqdq	\$0x00,$HK,$Xm
759	xorps		$Xl,$Xln
760	xorps		$Xh,$Xhn
761
762	lea	0x40($inp),$inp
763	sub	\$0x40,$len
764	jc	.Ltail4x
765
766	jmp	.Lmod4_loop
767.align	32
768.Lmod4_loop:
769	pclmulqdq	\$0x00,$Hkey4,$Xi
770	xorps		$Xm,$Xmn
771	 movdqu		0x30($inp),$Xl
772	 pshufb		$T3,$Xl
773	pclmulqdq	\$0x11,$Hkey4,$Xhi
774	xorps		$Xln,$Xi
775	 movdqu		0x20($inp),$Xln
776	 movdqa		$Xl,$Xh
777	pclmulqdq	\$0x10,$HK,$T1
778	 pshufd		\$0b01001110,$Xl,$Xm
779	xorps		$Xhn,$Xhi
780	 pxor		$Xl,$Xm
781	 pshufb		$T3,$Xln
782	movups		0x20($Htbl),$HK
783	xorps		$Xmn,$T1
784	 pclmulqdq	\$0x00,$Hkey,$Xl
785	 pshufd		\$0b01001110,$Xln,$Xmn
786
787	pxor		$Xi,$T1			# aggregated Karatsuba post-processing
788	 movdqa		$Xln,$Xhn
789	pxor		$Xhi,$T1		#
790	 pxor		$Xln,$Xmn
791	movdqa		$T1,$T2			#
792	 pclmulqdq	\$0x11,$Hkey,$Xh
793	pslldq		\$8,$T1
794	psrldq		\$8,$T2			#
795	pxor		$T1,$Xi
796	movdqa		.L7_mask(%rip),$T1
797	pxor		$T2,$Xhi		#
798	movq		%rax,$T2
799
800	pand		$Xi,$T1			# 1st phase
801	pshufb		$T1,$T2			#
802	pxor		$Xi,$T2			#
803	 pclmulqdq	\$0x00,$HK,$Xm
804	psllq		\$57,$T2		#
805	movdqa		$T2,$T1			#
806	pslldq		\$8,$T2
807	 pclmulqdq	\$0x00,$Hkey2,$Xln
808	psrldq		\$8,$T1			#
809	pxor		$T2,$Xi
810	pxor		$T1,$Xhi		#
811	movdqu		0($inp),$T1
812
813	movdqa		$Xi,$T2			# 2nd phase
814	psrlq		\$1,$Xi
815	 pclmulqdq	\$0x11,$Hkey2,$Xhn
816	 xorps		$Xl,$Xln
817	 movdqu		0x10($inp),$Xl
818	 pshufb		$T3,$Xl
819	 pclmulqdq	\$0x10,$HK,$Xmn
820	 xorps		$Xh,$Xhn
821	 movups		0x50($Htbl),$HK
822	pshufb		$T3,$T1
823	pxor		$T2,$Xhi		#
824	pxor		$Xi,$T2
825	psrlq		\$5,$Xi
826
827	 movdqa		$Xl,$Xh
828	 pxor		$Xm,$Xmn
829	 pshufd		\$0b01001110,$Xl,$Xm
830	pxor		$T2,$Xi			#
831	pxor		$T1,$Xhi
832	 pxor		$Xl,$Xm
833	 pclmulqdq	\$0x00,$Hkey3,$Xl
834	psrlq		\$1,$Xi			#
835	pxor		$Xhi,$Xi		#
836	movdqa		$Xi,$Xhi
837	 pclmulqdq	\$0x11,$Hkey3,$Xh
838	 xorps		$Xl,$Xln
839	pshufd		\$0b01001110,$Xi,$T1
840	pxor		$Xi,$T1
841
842	 pclmulqdq	\$0x00,$HK,$Xm
843	 xorps		$Xh,$Xhn
844
845	lea	0x40($inp),$inp
846	sub	\$0x40,$len
847	jnc	.Lmod4_loop
848
849.Ltail4x:
850	pclmulqdq	\$0x00,$Hkey4,$Xi
851	pclmulqdq	\$0x11,$Hkey4,$Xhi
852	pclmulqdq	\$0x10,$HK,$T1
853	xorps		$Xm,$Xmn
854	xorps		$Xln,$Xi
855	xorps		$Xhn,$Xhi
856	pxor		$Xi,$Xhi		# aggregated Karatsuba post-processing
857	pxor		$Xmn,$T1
858
859	pxor		$Xhi,$T1		#
860	pxor		$Xi,$Xhi
861
862	movdqa		$T1,$T2			#
863	psrldq		\$8,$T1
864	pslldq		\$8,$T2			#
865	pxor		$T1,$Xhi
866	pxor		$T2,$Xi			#
867___
868	&reduction_alg9($Xhi,$Xi);
869$code.=<<___;
870	add	\$0x40,$len
871	jz	.Ldone
872	movdqu	0x20($Htbl),$HK
873	sub	\$0x10,$len
874	jz	.Lodd_tail
875.Lskip4x:
876___
877}
878$code.=<<___;
879	#######
880	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
881	#	[(H*Ii+1) + (H*Xi+1)] mod P =
882	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
883	#
884	movdqu		($inp),$T1		# Ii
885	movdqu		16($inp),$Xln		# Ii+1
886	pshufb		$T3,$T1
887	pshufb		$T3,$Xln
888	pxor		$T1,$Xi			# Ii+Xi
889
890	movdqa		$Xln,$Xhn
891	pshufd		\$0b01001110,$Xln,$Xmn
892	pxor		$Xln,$Xmn
893	pclmulqdq	\$0x00,$Hkey,$Xln
894	pclmulqdq	\$0x11,$Hkey,$Xhn
895	pclmulqdq	\$0x00,$HK,$Xmn
896
897	lea		32($inp),$inp		# i+=2
898	nop
899	sub		\$0x20,$len
900	jbe		.Leven_tail
901	nop
902	jmp		.Lmod_loop
903
904.align	32
905.Lmod_loop:
906	movdqa		$Xi,$Xhi
907	movdqa		$Xmn,$T1
908	pshufd		\$0b01001110,$Xi,$Xmn	#
909	pxor		$Xi,$Xmn		#
910
911	pclmulqdq	\$0x00,$Hkey2,$Xi
912	pclmulqdq	\$0x11,$Hkey2,$Xhi
913	pclmulqdq	\$0x10,$HK,$Xmn
914
915	pxor		$Xln,$Xi		# (H*Ii+1) + H^2*(Ii+Xi)
916	pxor		$Xhn,$Xhi
917	  movdqu	($inp),$T2		# Ii
918	pxor		$Xi,$T1			# aggregated Karatsuba post-processing
919	  pshufb	$T3,$T2
920	  movdqu	16($inp),$Xln		# Ii+1
921
922	pxor		$Xhi,$T1
923	  pxor		$T2,$Xhi		# "Ii+Xi", consume early
924	pxor		$T1,$Xmn
925	 pshufb		$T3,$Xln
926	movdqa		$Xmn,$T1		#
927	psrldq		\$8,$T1
928	pslldq		\$8,$Xmn		#
929	pxor		$T1,$Xhi
930	pxor		$Xmn,$Xi		#
931
932	movdqa		$Xln,$Xhn		#
933
934	  movdqa	$Xi,$T2			# 1st phase
935	  movdqa	$Xi,$T1
936	  psllq		\$5,$Xi
937	  pxor		$Xi,$T1			#
938	pclmulqdq	\$0x00,$Hkey,$Xln	#######
939	  psllq		\$1,$Xi
940	  pxor		$T1,$Xi			#
941	  psllq		\$57,$Xi		#
942	  movdqa	$Xi,$T1			#
943	  pslldq	\$8,$Xi
944	  psrldq	\$8,$T1			#
945	  pxor		$T2,$Xi
946	pshufd		\$0b01001110,$Xhn,$Xmn
947	  pxor		$T1,$Xhi		#
948	pxor		$Xhn,$Xmn		#
949
950	  movdqa	$Xi,$T2			# 2nd phase
951	  psrlq		\$1,$Xi
952	pclmulqdq	\$0x11,$Hkey,$Xhn	#######
953	  pxor		$T2,$Xhi		#
954	  pxor		$Xi,$T2
955	  psrlq		\$5,$Xi
956	  pxor		$T2,$Xi			#
957	lea		32($inp),$inp
958	  psrlq		\$1,$Xi			#
959	pclmulqdq	\$0x00,$HK,$Xmn		#######
960	  pxor		$Xhi,$Xi		#
961
962	sub		\$0x20,$len
963	ja		.Lmod_loop
964
965.Leven_tail:
966	 movdqa		$Xi,$Xhi
967	 movdqa		$Xmn,$T1
968	 pshufd		\$0b01001110,$Xi,$Xmn	#
969	 pxor		$Xi,$Xmn		#
970
971	pclmulqdq	\$0x00,$Hkey2,$Xi
972	pclmulqdq	\$0x11,$Hkey2,$Xhi
973	pclmulqdq	\$0x10,$HK,$Xmn
974
975	pxor		$Xln,$Xi		# (H*Ii+1) + H^2*(Ii+Xi)
976	pxor		$Xhn,$Xhi
977	pxor		$Xi,$T1
978	pxor		$Xhi,$T1
979	pxor		$T1,$Xmn
980	movdqa		$Xmn,$T1		#
981	psrldq		\$8,$T1
982	pslldq		\$8,$Xmn		#
983	pxor		$T1,$Xhi
984	pxor		$Xmn,$Xi		#
985___
986	&reduction_alg9	($Xhi,$Xi);
987$code.=<<___;
988	test		$len,$len
989	jnz		.Ldone
990
991.Lodd_tail:
992	movdqu		($inp),$T1		# Ii
993	pshufb		$T3,$T1
994	pxor		$T1,$Xi			# Ii+Xi
995___
996	&clmul64x64_T2	($Xhi,$Xi,$Hkey,$HK);	# H*(Ii+Xi)
997	&reduction_alg9	($Xhi,$Xi);
998$code.=<<___;
999.Ldone:
1000	pshufb		$T3,$Xi
1001	movdqu		$Xi,($Xip)
1002___
1003$code.=<<___ if ($win64);
1004	movaps	(%rsp),%xmm6
1005	movaps	0x10(%rsp),%xmm7
1006	movaps	0x20(%rsp),%xmm8
1007	movaps	0x30(%rsp),%xmm9
1008	movaps	0x40(%rsp),%xmm10
1009	movaps	0x50(%rsp),%xmm11
1010	movaps	0x60(%rsp),%xmm12
1011	movaps	0x70(%rsp),%xmm13
1012	movaps	0x80(%rsp),%xmm14
1013	movaps	0x90(%rsp),%xmm15
1014	lea	0xa8(%rsp),%rsp
1015.LSEH_end_gcm_ghash_clmul:
1016___
1017$code.=<<___;
1018	ret
1019.cfi_endproc
1020.size	gcm_ghash_clmul,.-gcm_ghash_clmul
1021___
1022}
1023
1024$code.=<<___;
1025.globl	gcm_init_avx
1026.type	gcm_init_avx,\@abi-omnipotent
1027.align	32
1028gcm_init_avx:
1029.cfi_startproc
1030___
1031if ($avx) {
1032my ($Htbl,$Xip)=@_4args;
1033my $HK="%xmm6";
1034
1035$code.=<<___ if ($win64);
1036.LSEH_begin_gcm_init_avx:
1037	# I can't trust assembler to use specific encoding:-(
1038	.byte	0x48,0x83,0xec,0x18		#sub	$0x18,%rsp
1039	.byte	0x0f,0x29,0x34,0x24		#movaps	%xmm6,(%rsp)
1040___
1041$code.=<<___;
1042	vzeroupper
1043
1044	vmovdqu		($Xip),$Hkey
1045	vpshufd		\$0b01001110,$Hkey,$Hkey	# dword swap
1046
1047	# <<1 twist
1048	vpshufd		\$0b11111111,$Hkey,$T2	# broadcast uppermost dword
1049	vpsrlq		\$63,$Hkey,$T1
1050	vpsllq		\$1,$Hkey,$Hkey
1051	vpxor		$T3,$T3,$T3		#
1052	vpcmpgtd	$T2,$T3,$T3		# broadcast carry bit
1053	vpslldq		\$8,$T1,$T1
1054	vpor		$T1,$Hkey,$Hkey		# H<<=1
1055
1056	# magic reduction
1057	vpand		.L0x1c2_polynomial(%rip),$T3,$T3
1058	vpxor		$T3,$Hkey,$Hkey		# if(carry) H^=0x1c2_polynomial
1059
1060	vpunpckhqdq	$Hkey,$Hkey,$HK
1061	vmovdqa		$Hkey,$Xi
1062	vpxor		$Hkey,$HK,$HK
1063	mov		\$4,%r10		# up to H^8
1064	jmp		.Linit_start_avx
1065___
1066
1067sub clmul64x64_avx {
1068my ($Xhi,$Xi,$Hkey,$HK)=@_;
1069
1070if (!defined($HK)) {	$HK = $T2;
1071$code.=<<___;
1072	vpunpckhqdq	$Xi,$Xi,$T1
1073	vpunpckhqdq	$Hkey,$Hkey,$T2
1074	vpxor		$Xi,$T1,$T1		#
1075	vpxor		$Hkey,$T2,$T2
1076___
1077} else {
1078$code.=<<___;
1079	vpunpckhqdq	$Xi,$Xi,$T1
1080	vpxor		$Xi,$T1,$T1		#
1081___
1082}
1083$code.=<<___;
1084	vpclmulqdq	\$0x11,$Hkey,$Xi,$Xhi	#######
1085	vpclmulqdq	\$0x00,$Hkey,$Xi,$Xi	#######
1086	vpclmulqdq	\$0x00,$HK,$T1,$T1	#######
1087	vpxor		$Xi,$Xhi,$T2		#
1088	vpxor		$T2,$T1,$T1		#
1089
1090	vpslldq		\$8,$T1,$T2		#
1091	vpsrldq		\$8,$T1,$T1
1092	vpxor		$T2,$Xi,$Xi		#
1093	vpxor		$T1,$Xhi,$Xhi
1094___
1095}
1096
1097sub reduction_avx {
1098my ($Xhi,$Xi) = @_;
1099
1100$code.=<<___;
1101	vpsllq		\$57,$Xi,$T1		# 1st phase
1102	vpsllq		\$62,$Xi,$T2
1103	vpxor		$T1,$T2,$T2		#
1104	vpsllq		\$63,$Xi,$T1
1105	vpxor		$T1,$T2,$T2		#
1106	vpslldq		\$8,$T2,$T1		#
1107	vpsrldq		\$8,$T2,$T2
1108	vpxor		$T1,$Xi,$Xi		#
1109	vpxor		$T2,$Xhi,$Xhi
1110
1111	vpsrlq		\$1,$Xi,$T2		# 2nd phase
1112	vpxor		$Xi,$Xhi,$Xhi
1113	vpxor		$T2,$Xi,$Xi		#
1114	vpsrlq		\$5,$T2,$T2
1115	vpxor		$T2,$Xi,$Xi		#
1116	vpsrlq		\$1,$Xi,$Xi		#
1117	vpxor		$Xhi,$Xi,$Xi		#
1118___
1119}
1120
1121$code.=<<___;
1122.align	32
1123.Linit_loop_avx:
1124	vpalignr	\$8,$T1,$T2,$T3		# low part is H.lo^H.hi...
1125	vmovdqu		$T3,-0x10($Htbl)	# save Karatsuba "salt"
1126___
1127	&clmul64x64_avx	($Xhi,$Xi,$Hkey,$HK);	# calculate H^3,5,7
1128	&reduction_avx	($Xhi,$Xi);
1129$code.=<<___;
1130.Linit_start_avx:
1131	vmovdqa		$Xi,$T3
1132___
1133	&clmul64x64_avx	($Xhi,$Xi,$Hkey,$HK);	# calculate H^2,4,6,8
1134	&reduction_avx	($Xhi,$Xi);
1135$code.=<<___;
1136	vpshufd		\$0b01001110,$T3,$T1
1137	vpshufd		\$0b01001110,$Xi,$T2
1138	vpxor		$T3,$T1,$T1		# Karatsuba pre-processing
1139	vmovdqu		$T3,0x00($Htbl)		# save H^1,3,5,7
1140	vpxor		$Xi,$T2,$T2		# Karatsuba pre-processing
1141	vmovdqu		$Xi,0x10($Htbl)		# save H^2,4,6,8
1142	lea		0x30($Htbl),$Htbl
1143	sub		\$1,%r10
1144	jnz		.Linit_loop_avx
1145
1146	vpalignr	\$8,$T2,$T1,$T3		# last "salt" is flipped
1147	vmovdqu		$T3,-0x10($Htbl)
1148
1149	vzeroupper
1150___
1151$code.=<<___ if ($win64);
1152	movaps	(%rsp),%xmm6
1153	lea	0x18(%rsp),%rsp
1154.LSEH_end_gcm_init_avx:
1155___
1156$code.=<<___;
1157	ret
1158.cfi_endproc
1159.size	gcm_init_avx,.-gcm_init_avx
1160___
1161} else {
1162$code.=<<___;
1163	jmp	.L_init_clmul
1164.cfi_endproc
1165.size	gcm_init_avx,.-gcm_init_avx
1166___
1167}
1168
1169$code.=<<___;
1170.globl	gcm_gmult_avx
1171.type	gcm_gmult_avx,\@abi-omnipotent
1172.align	32
1173gcm_gmult_avx:
1174.cfi_startproc
1175	endbranch
1176	jmp	.L_gmult_clmul
1177.cfi_endproc
1178.size	gcm_gmult_avx,.-gcm_gmult_avx
1179___
1180
1181$code.=<<___;
1182.globl	gcm_ghash_avx
1183.type	gcm_ghash_avx,\@abi-omnipotent
1184.align	32
1185gcm_ghash_avx:
1186.cfi_startproc
1187	endbranch
1188___
1189if ($avx) {
1190my ($Xip,$Htbl,$inp,$len)=@_4args;
1191my ($Xlo,$Xhi,$Xmi,
1192    $Zlo,$Zhi,$Zmi,
1193    $Hkey,$HK,$T1,$T2,
1194    $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1195
1196$code.=<<___ if ($win64);
1197	lea	-0x88(%rsp),%rax
1198.LSEH_begin_gcm_ghash_avx:
1199	# I can't trust assembler to use specific encoding:-(
1200	.byte	0x48,0x8d,0x60,0xe0		#lea	-0x20(%rax),%rsp
1201	.byte	0x0f,0x29,0x70,0xe0		#movaps	%xmm6,-0x20(%rax)
1202	.byte	0x0f,0x29,0x78,0xf0		#movaps	%xmm7,-0x10(%rax)
1203	.byte	0x44,0x0f,0x29,0x00		#movaps	%xmm8,0(%rax)
1204	.byte	0x44,0x0f,0x29,0x48,0x10	#movaps	%xmm9,0x10(%rax)
1205	.byte	0x44,0x0f,0x29,0x50,0x20	#movaps	%xmm10,0x20(%rax)
1206	.byte	0x44,0x0f,0x29,0x58,0x30	#movaps	%xmm11,0x30(%rax)
1207	.byte	0x44,0x0f,0x29,0x60,0x40	#movaps	%xmm12,0x40(%rax)
1208	.byte	0x44,0x0f,0x29,0x68,0x50	#movaps	%xmm13,0x50(%rax)
1209	.byte	0x44,0x0f,0x29,0x70,0x60	#movaps	%xmm14,0x60(%rax)
1210	.byte	0x44,0x0f,0x29,0x78,0x70	#movaps	%xmm15,0x70(%rax)
1211___
1212$code.=<<___;
1213	vzeroupper
1214
1215	vmovdqu		($Xip),$Xi		# load $Xi
1216	lea		.L0x1c2_polynomial(%rip),%r10
1217	lea		0x40($Htbl),$Htbl	# size optimization
1218	vmovdqu		.Lbswap_mask(%rip),$bswap
1219	vpshufb		$bswap,$Xi,$Xi
1220	cmp		\$0x80,$len
1221	jb		.Lshort_avx
1222	sub		\$0x80,$len
1223
1224	vmovdqu		0x70($inp),$Ii		# I[7]
1225	vmovdqu		0x00-0x40($Htbl),$Hkey	# $Hkey^1
1226	vpshufb		$bswap,$Ii,$Ii
1227	vmovdqu		0x20-0x40($Htbl),$HK
1228
1229	vpunpckhqdq	$Ii,$Ii,$T2
1230	 vmovdqu	0x60($inp),$Ij		# I[6]
1231	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1232	vpxor		$Ii,$T2,$T2
1233	 vpshufb	$bswap,$Ij,$Ij
1234	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1235	 vmovdqu	0x10-0x40($Htbl),$Hkey	# $Hkey^2
1236	 vpunpckhqdq	$Ij,$Ij,$T1
1237	 vmovdqu	0x50($inp),$Ii		# I[5]
1238	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1239	 vpxor		$Ij,$T1,$T1
1240
1241	 vpshufb	$bswap,$Ii,$Ii
1242	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1243	 vpunpckhqdq	$Ii,$Ii,$T2
1244	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1245	 vmovdqu	0x30-0x40($Htbl),$Hkey	# $Hkey^3
1246	 vpxor		$Ii,$T2,$T2
1247	 vmovdqu	0x40($inp),$Ij		# I[4]
1248	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1249	 vmovdqu	0x50-0x40($Htbl),$HK
1250
1251	 vpshufb	$bswap,$Ij,$Ij
1252	vpxor		$Xlo,$Zlo,$Zlo
1253	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1254	vpxor		$Xhi,$Zhi,$Zhi
1255	 vpunpckhqdq	$Ij,$Ij,$T1
1256	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1257	 vmovdqu	0x40-0x40($Htbl),$Hkey	# $Hkey^4
1258	vpxor		$Xmi,$Zmi,$Zmi
1259	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1260	 vpxor		$Ij,$T1,$T1
1261
1262	 vmovdqu	0x30($inp),$Ii		# I[3]
1263	vpxor		$Zlo,$Xlo,$Xlo
1264	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1265	vpxor		$Zhi,$Xhi,$Xhi
1266	 vpshufb	$bswap,$Ii,$Ii
1267	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1268	 vmovdqu	0x60-0x40($Htbl),$Hkey	# $Hkey^5
1269	vpxor		$Zmi,$Xmi,$Xmi
1270	 vpunpckhqdq	$Ii,$Ii,$T2
1271	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1272	 vmovdqu	0x80-0x40($Htbl),$HK
1273	 vpxor		$Ii,$T2,$T2
1274
1275	 vmovdqu	0x20($inp),$Ij		# I[2]
1276	vpxor		$Xlo,$Zlo,$Zlo
1277	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1278	vpxor		$Xhi,$Zhi,$Zhi
1279	 vpshufb	$bswap,$Ij,$Ij
1280	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1281	 vmovdqu	0x70-0x40($Htbl),$Hkey	# $Hkey^6
1282	vpxor		$Xmi,$Zmi,$Zmi
1283	 vpunpckhqdq	$Ij,$Ij,$T1
1284	vpclmulqdq	\$0x00,$HK,$T2,$Xmi
1285	 vpxor		$Ij,$T1,$T1
1286
1287	 vmovdqu	0x10($inp),$Ii		# I[1]
1288	vpxor		$Zlo,$Xlo,$Xlo
1289	vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1290	vpxor		$Zhi,$Xhi,$Xhi
1291	 vpshufb	$bswap,$Ii,$Ii
1292	vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1293	 vmovdqu	0x90-0x40($Htbl),$Hkey	# $Hkey^7
1294	vpxor		$Zmi,$Xmi,$Xmi
1295	 vpunpckhqdq	$Ii,$Ii,$T2
1296	vpclmulqdq	\$0x10,$HK,$T1,$Zmi
1297	 vmovdqu	0xb0-0x40($Htbl),$HK
1298	 vpxor		$Ii,$T2,$T2
1299
1300	 vmovdqu	($inp),$Ij		# I[0]
1301	vpxor		$Xlo,$Zlo,$Zlo
1302	vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1303	vpxor		$Xhi,$Zhi,$Zhi
1304	 vpshufb	$bswap,$Ij,$Ij
1305	vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1306	 vmovdqu	0xa0-0x40($Htbl),$Hkey	# $Hkey^8
1307	vpxor		$Xmi,$Zmi,$Zmi
1308	vpclmulqdq	\$0x10,$HK,$T2,$Xmi
1309
1310	lea		0x80($inp),$inp
1311	cmp		\$0x80,$len
1312	jb		.Ltail_avx
1313
1314	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1315	sub		\$0x80,$len
1316	jmp		.Loop8x_avx
1317
1318.align	32
1319.Loop8x_avx:
1320	vpunpckhqdq	$Ij,$Ij,$T1
1321	 vmovdqu	0x70($inp),$Ii		# I[7]
1322	vpxor		$Xlo,$Zlo,$Zlo
1323	vpxor		$Ij,$T1,$T1
1324	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xi
1325	 vpshufb	$bswap,$Ii,$Ii
1326	vpxor		$Xhi,$Zhi,$Zhi
1327	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xo
1328	 vmovdqu	0x00-0x40($Htbl),$Hkey	# $Hkey^1
1329	 vpunpckhqdq	$Ii,$Ii,$T2
1330	vpxor		$Xmi,$Zmi,$Zmi
1331	vpclmulqdq	\$0x00,$HK,$T1,$Tred
1332	 vmovdqu	0x20-0x40($Htbl),$HK
1333	 vpxor		$Ii,$T2,$T2
1334
1335	  vmovdqu	0x60($inp),$Ij		# I[6]
1336	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1337	vpxor		$Zlo,$Xi,$Xi		# collect result
1338	  vpshufb	$bswap,$Ij,$Ij
1339	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1340	vxorps		$Zhi,$Xo,$Xo
1341	  vmovdqu	0x10-0x40($Htbl),$Hkey	# $Hkey^2
1342	 vpunpckhqdq	$Ij,$Ij,$T1
1343	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1344	vpxor		$Zmi,$Tred,$Tred
1345	 vxorps		$Ij,$T1,$T1
1346
1347	  vmovdqu	0x50($inp),$Ii		# I[5]
1348	vpxor		$Xi,$Tred,$Tred		# aggregated Karatsuba post-processing
1349	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1350	vpxor		$Xo,$Tred,$Tred
1351	vpslldq		\$8,$Tred,$T2
1352	 vpxor		$Xlo,$Zlo,$Zlo
1353	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1354	vpsrldq		\$8,$Tred,$Tred
1355	vpxor		$T2, $Xi, $Xi
1356	  vmovdqu	0x30-0x40($Htbl),$Hkey	# $Hkey^3
1357	  vpshufb	$bswap,$Ii,$Ii
1358	vxorps		$Tred,$Xo, $Xo
1359	 vpxor		$Xhi,$Zhi,$Zhi
1360	 vpunpckhqdq	$Ii,$Ii,$T2
1361	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1362	  vmovdqu	0x50-0x40($Htbl),$HK
1363	 vpxor		$Ii,$T2,$T2
1364	 vpxor		$Xmi,$Zmi,$Zmi
1365
1366	  vmovdqu	0x40($inp),$Ij		# I[4]
1367	vpalignr	\$8,$Xi,$Xi,$Tred	# 1st phase
1368	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1369	  vpshufb	$bswap,$Ij,$Ij
1370	 vpxor		$Zlo,$Xlo,$Xlo
1371	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1372	  vmovdqu	0x40-0x40($Htbl),$Hkey	# $Hkey^4
1373	 vpunpckhqdq	$Ij,$Ij,$T1
1374	 vpxor		$Zhi,$Xhi,$Xhi
1375	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1376	 vxorps		$Ij,$T1,$T1
1377	 vpxor		$Zmi,$Xmi,$Xmi
1378
1379	  vmovdqu	0x30($inp),$Ii		# I[3]
1380	vpclmulqdq	\$0x10,(%r10),$Xi,$Xi
1381	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1382	  vpshufb	$bswap,$Ii,$Ii
1383	 vpxor		$Xlo,$Zlo,$Zlo
1384	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1385	  vmovdqu	0x60-0x40($Htbl),$Hkey	# $Hkey^5
1386	 vpunpckhqdq	$Ii,$Ii,$T2
1387	 vpxor		$Xhi,$Zhi,$Zhi
1388	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1389	  vmovdqu	0x80-0x40($Htbl),$HK
1390	 vpxor		$Ii,$T2,$T2
1391	 vpxor		$Xmi,$Zmi,$Zmi
1392
1393	  vmovdqu	0x20($inp),$Ij		# I[2]
1394	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1395	  vpshufb	$bswap,$Ij,$Ij
1396	 vpxor		$Zlo,$Xlo,$Xlo
1397	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1398	  vmovdqu	0x70-0x40($Htbl),$Hkey	# $Hkey^6
1399	 vpunpckhqdq	$Ij,$Ij,$T1
1400	 vpxor		$Zhi,$Xhi,$Xhi
1401	 vpclmulqdq	\$0x00,$HK,  $T2,$Xmi
1402	 vpxor		$Ij,$T1,$T1
1403	 vpxor		$Zmi,$Xmi,$Xmi
1404	vxorps		$Tred,$Xi,$Xi
1405
1406	  vmovdqu	0x10($inp),$Ii		# I[1]
1407	vpalignr	\$8,$Xi,$Xi,$Tred	# 2nd phase
1408	 vpclmulqdq	\$0x00,$Hkey,$Ij,$Zlo
1409	  vpshufb	$bswap,$Ii,$Ii
1410	 vpxor		$Xlo,$Zlo,$Zlo
1411	 vpclmulqdq	\$0x11,$Hkey,$Ij,$Zhi
1412	  vmovdqu	0x90-0x40($Htbl),$Hkey	# $Hkey^7
1413	vpclmulqdq	\$0x10,(%r10),$Xi,$Xi
1414	vxorps		$Xo,$Tred,$Tred
1415	 vpunpckhqdq	$Ii,$Ii,$T2
1416	 vpxor		$Xhi,$Zhi,$Zhi
1417	 vpclmulqdq	\$0x10,$HK,  $T1,$Zmi
1418	  vmovdqu	0xb0-0x40($Htbl),$HK
1419	 vpxor		$Ii,$T2,$T2
1420	 vpxor		$Xmi,$Zmi,$Zmi
1421
1422	  vmovdqu	($inp),$Ij		# I[0]
1423	 vpclmulqdq	\$0x00,$Hkey,$Ii,$Xlo
1424	  vpshufb	$bswap,$Ij,$Ij
1425	 vpclmulqdq	\$0x11,$Hkey,$Ii,$Xhi
1426	  vmovdqu	0xa0-0x40($Htbl),$Hkey	# $Hkey^8
1427	vpxor		$Tred,$Ij,$Ij
1428	 vpclmulqdq	\$0x10,$HK,  $T2,$Xmi
1429	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1430
1431	lea		0x80($inp),$inp
1432	sub		\$0x80,$len
1433	jnc		.Loop8x_avx
1434
1435	add		\$0x80,$len
1436	jmp		.Ltail_no_xor_avx
1437
1438.align	32
1439.Lshort_avx:
1440	vmovdqu		-0x10($inp,$len),$Ii	# very last word
1441	lea		($inp,$len),$inp
1442	vmovdqu		0x00-0x40($Htbl),$Hkey	# $Hkey^1
1443	vmovdqu		0x20-0x40($Htbl),$HK
1444	vpshufb		$bswap,$Ii,$Ij
1445
1446	vmovdqa		$Xlo,$Zlo		# subtle way to zero $Zlo,
1447	vmovdqa		$Xhi,$Zhi		# $Zhi and
1448	vmovdqa		$Xmi,$Zmi		# $Zmi
1449	sub		\$0x10,$len
1450	jz		.Ltail_avx
1451
1452	vpunpckhqdq	$Ij,$Ij,$T1
1453	vpxor		$Xlo,$Zlo,$Zlo
1454	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1455	vpxor		$Ij,$T1,$T1
1456	 vmovdqu	-0x20($inp),$Ii
1457	vpxor		$Xhi,$Zhi,$Zhi
1458	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1459	vmovdqu		0x10-0x40($Htbl),$Hkey	# $Hkey^2
1460	 vpshufb	$bswap,$Ii,$Ij
1461	vpxor		$Xmi,$Zmi,$Zmi
1462	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1463	vpsrldq		\$8,$HK,$HK
1464	sub		\$0x10,$len
1465	jz		.Ltail_avx
1466
1467	vpunpckhqdq	$Ij,$Ij,$T1
1468	vpxor		$Xlo,$Zlo,$Zlo
1469	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1470	vpxor		$Ij,$T1,$T1
1471	 vmovdqu	-0x30($inp),$Ii
1472	vpxor		$Xhi,$Zhi,$Zhi
1473	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1474	vmovdqu		0x30-0x40($Htbl),$Hkey	# $Hkey^3
1475	 vpshufb	$bswap,$Ii,$Ij
1476	vpxor		$Xmi,$Zmi,$Zmi
1477	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1478	vmovdqu		0x50-0x40($Htbl),$HK
1479	sub		\$0x10,$len
1480	jz		.Ltail_avx
1481
1482	vpunpckhqdq	$Ij,$Ij,$T1
1483	vpxor		$Xlo,$Zlo,$Zlo
1484	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1485	vpxor		$Ij,$T1,$T1
1486	 vmovdqu	-0x40($inp),$Ii
1487	vpxor		$Xhi,$Zhi,$Zhi
1488	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1489	vmovdqu		0x40-0x40($Htbl),$Hkey	# $Hkey^4
1490	 vpshufb	$bswap,$Ii,$Ij
1491	vpxor		$Xmi,$Zmi,$Zmi
1492	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1493	vpsrldq		\$8,$HK,$HK
1494	sub		\$0x10,$len
1495	jz		.Ltail_avx
1496
1497	vpunpckhqdq	$Ij,$Ij,$T1
1498	vpxor		$Xlo,$Zlo,$Zlo
1499	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1500	vpxor		$Ij,$T1,$T1
1501	 vmovdqu	-0x50($inp),$Ii
1502	vpxor		$Xhi,$Zhi,$Zhi
1503	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1504	vmovdqu		0x60-0x40($Htbl),$Hkey	# $Hkey^5
1505	 vpshufb	$bswap,$Ii,$Ij
1506	vpxor		$Xmi,$Zmi,$Zmi
1507	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1508	vmovdqu		0x80-0x40($Htbl),$HK
1509	sub		\$0x10,$len
1510	jz		.Ltail_avx
1511
1512	vpunpckhqdq	$Ij,$Ij,$T1
1513	vpxor		$Xlo,$Zlo,$Zlo
1514	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1515	vpxor		$Ij,$T1,$T1
1516	 vmovdqu	-0x60($inp),$Ii
1517	vpxor		$Xhi,$Zhi,$Zhi
1518	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1519	vmovdqu		0x70-0x40($Htbl),$Hkey	# $Hkey^6
1520	 vpshufb	$bswap,$Ii,$Ij
1521	vpxor		$Xmi,$Zmi,$Zmi
1522	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1523	vpsrldq		\$8,$HK,$HK
1524	sub		\$0x10,$len
1525	jz		.Ltail_avx
1526
1527	vpunpckhqdq	$Ij,$Ij,$T1
1528	vpxor		$Xlo,$Zlo,$Zlo
1529	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1530	vpxor		$Ij,$T1,$T1
1531	 vmovdqu	-0x70($inp),$Ii
1532	vpxor		$Xhi,$Zhi,$Zhi
1533	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1534	vmovdqu		0x90-0x40($Htbl),$Hkey	# $Hkey^7
1535	 vpshufb	$bswap,$Ii,$Ij
1536	vpxor		$Xmi,$Zmi,$Zmi
1537	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1538	vmovq		0xb8-0x40($Htbl),$HK
1539	sub		\$0x10,$len
1540	jmp		.Ltail_avx
1541
1542.align	32
1543.Ltail_avx:
1544	vpxor		$Xi,$Ij,$Ij		# accumulate $Xi
1545.Ltail_no_xor_avx:
1546	vpunpckhqdq	$Ij,$Ij,$T1
1547	vpxor		$Xlo,$Zlo,$Zlo
1548	vpclmulqdq	\$0x00,$Hkey,$Ij,$Xlo
1549	vpxor		$Ij,$T1,$T1
1550	vpxor		$Xhi,$Zhi,$Zhi
1551	vpclmulqdq	\$0x11,$Hkey,$Ij,$Xhi
1552	vpxor		$Xmi,$Zmi,$Zmi
1553	vpclmulqdq	\$0x00,$HK,$T1,$Xmi
1554
1555	vmovdqu		(%r10),$Tred
1556
1557	vpxor		$Xlo,$Zlo,$Xi
1558	vpxor		$Xhi,$Zhi,$Xo
1559	vpxor		$Xmi,$Zmi,$Zmi
1560
1561	vpxor		$Xi, $Zmi,$Zmi		# aggregated Karatsuba post-processing
1562	vpxor		$Xo, $Zmi,$Zmi
1563	vpslldq		\$8, $Zmi,$T2
1564	vpsrldq		\$8, $Zmi,$Zmi
1565	vpxor		$T2, $Xi, $Xi
1566	vpxor		$Zmi,$Xo, $Xo
1567
1568	vpclmulqdq	\$0x10,$Tred,$Xi,$T2	# 1st phase
1569	vpalignr	\$8,$Xi,$Xi,$Xi
1570	vpxor		$T2,$Xi,$Xi
1571
1572	vpclmulqdq	\$0x10,$Tred,$Xi,$T2	# 2nd phase
1573	vpalignr	\$8,$Xi,$Xi,$Xi
1574	vpxor		$Xo,$Xi,$Xi
1575	vpxor		$T2,$Xi,$Xi
1576
1577	cmp		\$0,$len
1578	jne		.Lshort_avx
1579
1580	vpshufb		$bswap,$Xi,$Xi
1581	vmovdqu		$Xi,($Xip)
1582	vzeroupper
1583___
1584$code.=<<___ if ($win64);
1585	movaps	(%rsp),%xmm6
1586	movaps	0x10(%rsp),%xmm7
1587	movaps	0x20(%rsp),%xmm8
1588	movaps	0x30(%rsp),%xmm9
1589	movaps	0x40(%rsp),%xmm10
1590	movaps	0x50(%rsp),%xmm11
1591	movaps	0x60(%rsp),%xmm12
1592	movaps	0x70(%rsp),%xmm13
1593	movaps	0x80(%rsp),%xmm14
1594	movaps	0x90(%rsp),%xmm15
1595	lea	0xa8(%rsp),%rsp
1596.LSEH_end_gcm_ghash_avx:
1597___
1598$code.=<<___;
1599	ret
1600.cfi_endproc
1601.size	gcm_ghash_avx,.-gcm_ghash_avx
1602___
1603} else {
1604$code.=<<___;
1605	jmp	.L_ghash_clmul
1606.cfi_endproc
1607.size	gcm_ghash_avx,.-gcm_ghash_avx
1608___
1609}
1610
1611$code.=<<___;
1612.align	64
1613.Lbswap_mask:
1614	.byte	15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1615.L0x1c2_polynomial:
1616	.byte	1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1617.L7_mask:
1618	.long	7,0,7,0
1619.L7_mask_poly:
1620	.long	7,0,`0xE1<<1`,0
1621.align	64
1622.type	.Lrem_4bit,\@object
1623.Lrem_4bit:
1624	.long	0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1625	.long	0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1626	.long	0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1627	.long	0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1628.type	.Lrem_8bit,\@object
1629.Lrem_8bit:
1630	.value	0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1631	.value	0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1632	.value	0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1633	.value	0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1634	.value	0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1635	.value	0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1636	.value	0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1637	.value	0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1638	.value	0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1639	.value	0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1640	.value	0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1641	.value	0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1642	.value	0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1643	.value	0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1644	.value	0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1645	.value	0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1646	.value	0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1647	.value	0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1648	.value	0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1649	.value	0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1650	.value	0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1651	.value	0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1652	.value	0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1653	.value	0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1654	.value	0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1655	.value	0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1656	.value	0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1657	.value	0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1658	.value	0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1659	.value	0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1660	.value	0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1661	.value	0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1662
1663.asciz	"GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1664.align	64
1665___
1666
1667# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1668#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
1669if ($win64) {
1670$rec="%rcx";
1671$frame="%rdx";
1672$context="%r8";
1673$disp="%r9";
1674
1675$code.=<<___;
1676.extern	__imp_RtlVirtualUnwind
1677.type	se_handler,\@abi-omnipotent
1678.align	16
1679se_handler:
1680	push	%rsi
1681	push	%rdi
1682	push	%rbx
1683	push	%rbp
1684	push	%r12
1685	push	%r13
1686	push	%r14
1687	push	%r15
1688	pushfq
1689	sub	\$64,%rsp
1690
1691	mov	120($context),%rax	# pull context->Rax
1692	mov	248($context),%rbx	# pull context->Rip
1693
1694	mov	8($disp),%rsi		# disp->ImageBase
1695	mov	56($disp),%r11		# disp->HandlerData
1696
1697	mov	0(%r11),%r10d		# HandlerData[0]
1698	lea	(%rsi,%r10),%r10	# prologue label
1699	cmp	%r10,%rbx		# context->Rip<prologue label
1700	jb	.Lin_prologue
1701
1702	mov	152($context),%rax	# pull context->Rsp
1703
1704	mov	4(%r11),%r10d		# HandlerData[1]
1705	lea	(%rsi,%r10),%r10	# epilogue label
1706	cmp	%r10,%rbx		# context->Rip>=epilogue label
1707	jae	.Lin_prologue
1708
1709	lea	48+280(%rax),%rax	# adjust "rsp"
1710
1711	mov	-8(%rax),%rbx
1712	mov	-16(%rax),%rbp
1713	mov	-24(%rax),%r12
1714	mov	-32(%rax),%r13
1715	mov	-40(%rax),%r14
1716	mov	-48(%rax),%r15
1717	mov	%rbx,144($context)	# restore context->Rbx
1718	mov	%rbp,160($context)	# restore context->Rbp
1719	mov	%r12,216($context)	# restore context->R12
1720	mov	%r13,224($context)	# restore context->R13
1721	mov	%r14,232($context)	# restore context->R14
1722	mov	%r15,240($context)	# restore context->R15
1723
1724.Lin_prologue:
1725	mov	8(%rax),%rdi
1726	mov	16(%rax),%rsi
1727	mov	%rax,152($context)	# restore context->Rsp
1728	mov	%rsi,168($context)	# restore context->Rsi
1729	mov	%rdi,176($context)	# restore context->Rdi
1730
1731	mov	40($disp),%rdi		# disp->ContextRecord
1732	mov	$context,%rsi		# context
1733	mov	\$`1232/8`,%ecx		# sizeof(CONTEXT)
1734	.long	0xa548f3fc		# cld; rep movsq
1735
1736	mov	$disp,%rsi
1737	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
1738	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
1739	mov	0(%rsi),%r8		# arg3, disp->ControlPc
1740	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
1741	mov	40(%rsi),%r10		# disp->ContextRecord
1742	lea	56(%rsi),%r11		# &disp->HandlerData
1743	lea	24(%rsi),%r12		# &disp->EstablisherFrame
1744	mov	%r10,32(%rsp)		# arg5
1745	mov	%r11,40(%rsp)		# arg6
1746	mov	%r12,48(%rsp)		# arg7
1747	mov	%rcx,56(%rsp)		# arg8, (NULL)
1748	call	*__imp_RtlVirtualUnwind(%rip)
1749
1750	mov	\$1,%eax		# ExceptionContinueSearch
1751	add	\$64,%rsp
1752	popfq
1753	pop	%r15
1754	pop	%r14
1755	pop	%r13
1756	pop	%r12
1757	pop	%rbp
1758	pop	%rbx
1759	pop	%rdi
1760	pop	%rsi
1761	ret
1762.size	se_handler,.-se_handler
1763
1764.section	.pdata
1765.align	4
1766	.rva	.LSEH_begin_gcm_gmult_4bit
1767	.rva	.LSEH_end_gcm_gmult_4bit
1768	.rva	.LSEH_info_gcm_gmult_4bit
1769
1770	.rva	.LSEH_begin_gcm_ghash_4bit
1771	.rva	.LSEH_end_gcm_ghash_4bit
1772	.rva	.LSEH_info_gcm_ghash_4bit
1773
1774	.rva	.LSEH_begin_gcm_init_clmul
1775	.rva	.LSEH_end_gcm_init_clmul
1776	.rva	.LSEH_info_gcm_init_clmul
1777
1778	.rva	.LSEH_begin_gcm_ghash_clmul
1779	.rva	.LSEH_end_gcm_ghash_clmul
1780	.rva	.LSEH_info_gcm_ghash_clmul
1781___
1782$code.=<<___	if ($avx);
1783	.rva	.LSEH_begin_gcm_init_avx
1784	.rva	.LSEH_end_gcm_init_avx
1785	.rva	.LSEH_info_gcm_init_clmul
1786
1787	.rva	.LSEH_begin_gcm_ghash_avx
1788	.rva	.LSEH_end_gcm_ghash_avx
1789	.rva	.LSEH_info_gcm_ghash_clmul
1790___
1791$code.=<<___;
1792.section	.xdata
1793.align	8
1794.LSEH_info_gcm_gmult_4bit:
1795	.byte	9,0,0,0
1796	.rva	se_handler
1797	.rva	.Lgmult_prologue,.Lgmult_epilogue	# HandlerData
1798.LSEH_info_gcm_ghash_4bit:
1799	.byte	9,0,0,0
1800	.rva	se_handler
1801	.rva	.Lghash_prologue,.Lghash_epilogue	# HandlerData
1802.LSEH_info_gcm_init_clmul:
1803	.byte	0x01,0x08,0x03,0x00
1804	.byte	0x08,0x68,0x00,0x00	#movaps	0x00(rsp),xmm6
1805	.byte	0x04,0x22,0x00,0x00	#sub	rsp,0x18
1806.LSEH_info_gcm_ghash_clmul:
1807	.byte	0x01,0x33,0x16,0x00
1808	.byte	0x33,0xf8,0x09,0x00	#movaps 0x90(rsp),xmm15
1809	.byte	0x2e,0xe8,0x08,0x00	#movaps 0x80(rsp),xmm14
1810	.byte	0x29,0xd8,0x07,0x00	#movaps 0x70(rsp),xmm13
1811	.byte	0x24,0xc8,0x06,0x00	#movaps 0x60(rsp),xmm12
1812	.byte	0x1f,0xb8,0x05,0x00	#movaps 0x50(rsp),xmm11
1813	.byte	0x1a,0xa8,0x04,0x00	#movaps 0x40(rsp),xmm10
1814	.byte	0x15,0x98,0x03,0x00	#movaps 0x30(rsp),xmm9
1815	.byte	0x10,0x88,0x02,0x00	#movaps 0x20(rsp),xmm8
1816	.byte	0x0c,0x78,0x01,0x00	#movaps 0x10(rsp),xmm7
1817	.byte	0x08,0x68,0x00,0x00	#movaps 0x00(rsp),xmm6
1818	.byte	0x04,0x01,0x15,0x00	#sub	rsp,0xa8
1819___
1820}
1821
1822$code =~ s/\`([^\`]*)\`/eval($1)/gem;
1823
1824print $code;
1825
1826close STDOUT or die "error closing STDOUT: $!";
1827