• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#!/usr/bin/env perl
2
3# ====================================================================
4# [Re]written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9
10# "[Re]written" was achieved in two major overhauls. In 2004 BODY_*
11# functions were re-implemented to address P4 performance issue [see
12# commentary below], and in 2006 the rest was rewritten in order to
13# gain freedom to liberate licensing terms.
14
15# January, September 2004.
16#
17# It was noted that Intel IA-32 C compiler generates code which
18# performs ~30% *faster* on P4 CPU than original *hand-coded*
19# SHA1 assembler implementation. To address this problem (and
20# prove that humans are still better than machines:-), the
21# original code was overhauled, which resulted in following
22# performance changes:
23#
24#		compared with original	compared with Intel cc
25#		assembler impl.		generated code
26# Pentium	-16%			+48%
27# PIII/AMD	+8%			+16%
28# P4		+85%(!)			+45%
29#
30# As you can see Pentium came out as looser:-( Yet I reckoned that
31# improvement on P4 outweights the loss and incorporate this
32# re-tuned code to 0.9.7 and later.
33# ----------------------------------------------------------------
34#					<appro@fy.chalmers.se>
35
36# August 2009.
37#
38# George Spelvin has tipped that F_40_59(b,c,d) can be rewritten as
39# '(c&d) + (b&(c^d))', which allows to accumulate partial results
40# and lighten "pressure" on scratch registers. This resulted in
41# >12% performance improvement on contemporary AMD cores (with no
42# degradation on other CPUs:-). Also, the code was revised to maximize
43# "distance" between instructions producing input to 'lea' instruction
44# and the 'lea' instruction itself, which is essential for Intel Atom
45# core and resulted in ~15% improvement.
46
47# October 2010.
48#
49# Add SSSE3, Supplemental[!] SSE3, implementation. The idea behind it
50# is to offload message schedule denoted by Wt in NIST specification,
51# or Xupdate in OpenSSL source, to SIMD unit. The idea is not novel,
52# and in SSE2 context was first explored by Dean Gaudet in 2004, see
53# http://arctic.org/~dean/crypto/sha1.html. Since then several things
54# have changed that made it interesting again:
55#
56# a) XMM units became faster and wider;
57# b) instruction set became more versatile;
58# c) an important observation was made by Max Locktykhin, which made
59#    it possible to reduce amount of instructions required to perform
60#    the operation in question, for further details see
61#    http://software.intel.com/en-us/articles/improving-the-performance-of-the-secure-hash-algorithm-1/.
62
63# April 2011.
64#
65# Add AVX code path, probably most controversial... The thing is that
66# switch to AVX alone improves performance by as little as 4% in
67# comparison to SSSE3 code path. But below result doesn't look like
68# 4% improvement... Trouble is that Sandy Bridge decodes 'ro[rl]' as
69# pair of µ-ops, and it's the additional µ-ops, two per round, that
70# make it run slower than Core2 and Westmere. But 'sh[rl]d' is decoded
71# as single µ-op by Sandy Bridge and it's replacing 'ro[rl]' with
72# equivalent 'sh[rl]d' that is responsible for the impressive 5.1
73# cycles per processed byte. But 'sh[rl]d' is not something that used
74# to be fast, nor does it appear to be fast in upcoming Bulldozer
75# [according to its optimization manual]. Which is why AVX code path
76# is guarded by *both* AVX and synthetic bit denoting Intel CPUs.
77# One can argue that it's unfair to AMD, but without 'sh[rl]d' it
78# makes no sense to keep the AVX code path. If somebody feels that
79# strongly, it's probably more appropriate to discuss possibility of
80# using vector rotate XOP on AMD...
81
82######################################################################
83# Current performance is summarized in following table. Numbers are
84# CPU clock cycles spent to process single byte (less is better).
85#
86#		x86		SSSE3		AVX
87# Pentium	15.7		-
88# PIII		11.5		-
89# P4		10.6		-
90# AMD K8	7.1		-
91# Core2		7.3		6.1/+20%	-
92# Atom		12.5		9.5(*)/+32%	-
93# Westmere	7.3		5.6/+30%	-
94# Sandy Bridge	8.8		6.2/+40%	5.1(**)/+70%
95#
96# (*)	Loop is 1056 instructions long and expected result is ~8.25.
97#	It remains mystery [to me] why ILP is limited to 1.7.
98#
99# (**)	As per above comment, the result is for AVX *plus* sh[rl]d.
100
101$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
102push(@INC,"${dir}","${dir}../../perlasm");
103require "x86asm.pl";
104
105&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
106
107$xmm=$ymm=0;
108for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
109
110$ymm=1 if ($xmm &&
111		`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
112			=~ /GNU assembler version ([2-9]\.[0-9]+)/ &&
113		$1>=2.19);	# first version supporting AVX
114
115$ymm=1 if ($xmm && !$ymm && $ARGV[0] eq "win32n" &&
116		`nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/ &&
117		$1>=2.03);	# first version supporting AVX
118
119&external_label("OPENSSL_ia32cap_P") if ($xmm);
120
121
122$A="eax";
123$B="ebx";
124$C="ecx";
125$D="edx";
126$E="edi";
127$T="esi";
128$tmp1="ebp";
129
130@V=($A,$B,$C,$D,$E,$T);
131
132$alt=0;	# 1 denotes alternative IALU implementation, which performs
133	# 8% *worse* on P4, same on Westmere and Atom, 2% better on
134	# Sandy Bridge...
135
136sub BODY_00_15
137	{
138	local($n,$a,$b,$c,$d,$e,$f)=@_;
139
140	&comment("00_15 $n");
141
142	&mov($f,$c);			# f to hold F_00_19(b,c,d)
143	 if ($n==0)  { &mov($tmp1,$a); }
144	 else        { &mov($a,$tmp1); }
145	&rotl($tmp1,5);			# tmp1=ROTATE(a,5)
146	 &xor($f,$d);
147	&add($tmp1,$e);			# tmp1+=e;
148	 &mov($e,&swtmp($n%16));	# e becomes volatile and is loaded
149	 				# with xi, also note that e becomes
150					# f in next round...
151	&and($f,$b);
152	&rotr($b,2);			# b=ROTATE(b,30)
153	 &xor($f,$d);			# f holds F_00_19(b,c,d)
154	&lea($tmp1,&DWP(0x5a827999,$tmp1,$e));	# tmp1+=K_00_19+xi
155
156	if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
157		      &add($f,$tmp1); }	# f+=tmp1
158	else        { &add($tmp1,$f); }	# f becomes a in next round
159	&mov($tmp1,$a)			if ($alt && $n==15);
160	}
161
162sub BODY_16_19
163	{
164	local($n,$a,$b,$c,$d,$e,$f)=@_;
165
166	&comment("16_19 $n");
167
168if ($alt) {
169	&xor($c,$d);
170	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
171	&and($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d), b&=c^d
172	 &xor($f,&swtmp(($n+8)%16));
173	&xor($tmp1,$d);			# tmp1=F_00_19(b,c,d)
174	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
175	&rotl($f,1);			# f=ROTATE(f,1)
176	 &add($e,$tmp1);		# e+=F_00_19(b,c,d)
177	&xor($c,$d);			# restore $c
178	 &mov($tmp1,$a);		# b in next round
179	&rotr($b,$n==16?2:7);		# b=ROTATE(b,30)
180	 &mov(&swtmp($n%16),$f);	# xi=f
181	&rotl($a,5);			# ROTATE(a,5)
182	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
183	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
184	 &add($f,$a);			# f+=ROTATE(a,5)
185} else {
186	&mov($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d)
187	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
188	&xor($tmp1,$d);
189	 &xor($f,&swtmp(($n+8)%16));
190	&and($tmp1,$b);
191	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
192	&rotl($f,1);			# f=ROTATE(f,1)
193	 &xor($tmp1,$d);		# tmp1=F_00_19(b,c,d)
194	&add($e,$tmp1);			# e+=F_00_19(b,c,d)
195	 &mov($tmp1,$a);
196	&rotr($b,2);			# b=ROTATE(b,30)
197	 &mov(&swtmp($n%16),$f);	# xi=f
198	&rotl($tmp1,5);			# ROTATE(a,5)
199	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
200	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
201	 &add($f,$tmp1);		# f+=ROTATE(a,5)
202}
203	}
204
205sub BODY_20_39
206	{
207	local($n,$a,$b,$c,$d,$e,$f)=@_;
208	local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
209
210	&comment("20_39 $n");
211
212if ($alt) {
213	&xor($tmp1,$c);			# tmp1 to hold F_20_39(b,c,d), b^=c
214	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
215	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
216	 &xor($f,&swtmp(($n+8)%16));
217	&add($e,$tmp1);			# e+=F_20_39(b,c,d)
218	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
219	&rotl($f,1);			# f=ROTATE(f,1)
220	 &mov($tmp1,$a);		# b in next round
221	&rotr($b,7);			# b=ROTATE(b,30)
222	 &mov(&swtmp($n%16),$f)		if($n<77);# xi=f
223	&rotl($a,5);			# ROTATE(a,5)
224	 &xor($b,$c)			if($n==39);# warm up for BODY_40_59
225	&and($tmp1,$b)			if($n==39);
226	 &lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
227	&mov($e,&swtmp(($n+1)%16))	if($n<79);# pre-fetch f for next round
228	 &add($f,$a);			# f+=ROTATE(a,5)
229	&rotr($a,5)			if ($n==79);
230} else {
231	&mov($tmp1,$b);			# tmp1 to hold F_20_39(b,c,d)
232	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
233	&xor($tmp1,$c);
234	 &xor($f,&swtmp(($n+8)%16));
235	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
236	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
237	&rotl($f,1);			# f=ROTATE(f,1)
238	 &add($e,$tmp1);		# e+=F_20_39(b,c,d)
239	&rotr($b,2);			# b=ROTATE(b,30)
240	 &mov($tmp1,$a);
241	&rotl($tmp1,5);			# ROTATE(a,5)
242	 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
243	&lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
244	 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
245	&add($f,$tmp1);			# f+=ROTATE(a,5)
246}
247	}
248
249sub BODY_40_59
250	{
251	local($n,$a,$b,$c,$d,$e,$f)=@_;
252
253	&comment("40_59 $n");
254
255if ($alt) {
256	&add($e,$tmp1);			# e+=b&(c^d)
257	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
258	&mov($tmp1,$d);
259	 &xor($f,&swtmp(($n+8)%16));
260	&xor($c,$d);			# restore $c
261	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
262	&rotl($f,1);			# f=ROTATE(f,1)
263	 &and($tmp1,$c);
264	&rotr($b,7);			# b=ROTATE(b,30)
265	 &add($e,$tmp1);		# e+=c&d
266	&mov($tmp1,$a);			# b in next round
267	 &mov(&swtmp($n%16),$f);	# xi=f
268	&rotl($a,5);			# ROTATE(a,5)
269	 &xor($b,$c)			if ($n<59);
270	&and($tmp1,$b)			if ($n<59);# tmp1 to hold F_40_59(b,c,d)
271	 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
272	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
273	 &add($f,$a);			# f+=ROTATE(a,5)
274} else {
275	&mov($tmp1,$c);			# tmp1 to hold F_40_59(b,c,d)
276	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
277	&xor($tmp1,$d);
278	 &xor($f,&swtmp(($n+8)%16));
279	&and($tmp1,$b);
280	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
281	&rotl($f,1);			# f=ROTATE(f,1)
282	 &add($tmp1,$e);		# b&(c^d)+=e
283	&rotr($b,2);			# b=ROTATE(b,30)
284	 &mov($e,$a);			# e becomes volatile
285	&rotl($e,5);			# ROTATE(a,5)
286	 &mov(&swtmp($n%16),$f);	# xi=f
287	&lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
288	 &mov($tmp1,$c);
289	&add($f,$e);			# f+=ROTATE(a,5)
290	 &and($tmp1,$d);
291	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
292	 &add($f,$tmp1);		# f+=c&d
293}
294	}
295
296&function_begin("sha1_block_data_order");
297if ($xmm) {
298  &static_label("ssse3_shortcut");
299  &static_label("avx_shortcut")		if ($ymm);
300  &static_label("K_XX_XX");
301
302	&call	(&label("pic_point"));	# make it PIC!
303  &set_label("pic_point");
304	&blindpop($tmp1);
305	&picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
306	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
307
308	&mov	($A,&DWP(0,$T));
309	&mov	($D,&DWP(4,$T));
310	&test	($D,1<<9);		# check SSSE3 bit
311	&jz	(&label("x86"));
312	&test	($A,1<<24);		# check FXSR bit
313	&jz	(&label("x86"));
314	if ($ymm) {
315		&and	($D,1<<28);		# mask AVX bit
316		&and	($A,1<<30);		# mask "Intel CPU" bit
317		&or	($A,$D);
318		&cmp	($A,1<<28|1<<30);
319		&je	(&label("avx_shortcut"));
320	}
321	&jmp	(&label("ssse3_shortcut"));
322  &set_label("x86",16);
323}
324	&mov($tmp1,&wparam(0));	# SHA_CTX *c
325	&mov($T,&wparam(1));	# const void *input
326	&mov($A,&wparam(2));	# size_t num
327	&stack_push(16+3);	# allocate X[16]
328	&shl($A,6);
329	&add($A,$T);
330	&mov(&wparam(2),$A);	# pointer beyond the end of input
331	&mov($E,&DWP(16,$tmp1));# pre-load E
332	&jmp(&label("loop"));
333
334&set_label("loop",16);
335
336	# copy input chunk to X, but reversing byte order!
337	for ($i=0; $i<16; $i+=4)
338		{
339		&mov($A,&DWP(4*($i+0),$T));
340		&mov($B,&DWP(4*($i+1),$T));
341		&mov($C,&DWP(4*($i+2),$T));
342		&mov($D,&DWP(4*($i+3),$T));
343		&bswap($A);
344		&bswap($B);
345		&bswap($C);
346		&bswap($D);
347		&mov(&swtmp($i+0),$A);
348		&mov(&swtmp($i+1),$B);
349		&mov(&swtmp($i+2),$C);
350		&mov(&swtmp($i+3),$D);
351		}
352	&mov(&wparam(1),$T);	# redundant in 1st spin
353
354	&mov($A,&DWP(0,$tmp1));	# load SHA_CTX
355	&mov($B,&DWP(4,$tmp1));
356	&mov($C,&DWP(8,$tmp1));
357	&mov($D,&DWP(12,$tmp1));
358	# E is pre-loaded
359
360	for($i=0;$i<16;$i++)	{ &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
361	for(;$i<20;$i++)	{ &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
362	for(;$i<40;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
363	for(;$i<60;$i++)	{ &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
364	for(;$i<80;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
365
366	(($V[5] eq $D) and ($V[0] eq $E)) or die;	# double-check
367
368	&mov($tmp1,&wparam(0));	# re-load SHA_CTX*
369	&mov($D,&wparam(1));	# D is last "T" and is discarded
370
371	&add($E,&DWP(0,$tmp1));	# E is last "A"...
372	&add($T,&DWP(4,$tmp1));
373	&add($A,&DWP(8,$tmp1));
374	&add($B,&DWP(12,$tmp1));
375	&add($C,&DWP(16,$tmp1));
376
377	&mov(&DWP(0,$tmp1),$E);	# update SHA_CTX
378	 &add($D,64);		# advance input pointer
379	&mov(&DWP(4,$tmp1),$T);
380	 &cmp($D,&wparam(2));	# have we reached the end yet?
381	&mov(&DWP(8,$tmp1),$A);
382	 &mov($E,$C);		# C is last "E" which needs to be "pre-loaded"
383	&mov(&DWP(12,$tmp1),$B);
384	 &mov($T,$D);		# input pointer
385	&mov(&DWP(16,$tmp1),$C);
386	&jb(&label("loop"));
387
388	&stack_pop(16+3);
389&function_end("sha1_block_data_order");
390
391if ($xmm) {
392######################################################################
393# The SSSE3 implementation.
394#
395# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
396# 32 elements of the message schedule or Xupdate outputs. First 4
397# quadruples are simply byte-swapped input, next 4 are calculated
398# according to method originally suggested by Dean Gaudet (modulo
399# being implemented in SSSE3). Once 8 quadruples or 32 elements are
400# collected, it switches to routine proposed by Max Locktyukhin.
401#
402# Calculations inevitably require temporary reqisters, and there are
403# no %xmm registers left to spare. For this reason part of the ring
404# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
405# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
406# X[-5], and X[4] - X[-4]...
407#
408# Another notable optimization is aggressive stack frame compression
409# aiming to minimize amount of 9-byte instructions...
410#
411# Yet another notable optimization is "jumping" $B variable. It means
412# that there is no register permanently allocated for $B value. This
413# allowed to eliminate one instruction from body_20_39...
414#
415my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
416my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
417my @V=($A,$B,$C,$D,$E);
418my $j=0;			# hash round
419my @T=($T,$tmp1);
420my $inp;
421
422my $_rol=sub { &rol(@_) };
423my $_ror=sub { &ror(@_) };
424
425&function_begin("_sha1_block_data_order_ssse3");
426	&call	(&label("pic_point"));	# make it PIC!
427	&set_label("pic_point");
428	&blindpop($tmp1);
429	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
430&set_label("ssse3_shortcut");
431
432	&movdqa	(@X[3],&QWP(0,$tmp1));		# K_00_19
433	&movdqa	(@X[4],&QWP(16,$tmp1));		# K_20_39
434	&movdqa	(@X[5],&QWP(32,$tmp1));		# K_40_59
435	&movdqa	(@X[6],&QWP(48,$tmp1));		# K_60_79
436	&movdqa	(@X[2],&QWP(64,$tmp1));		# pbswap mask
437
438	&mov	($E,&wparam(0));		# load argument block
439	&mov	($inp=@T[1],&wparam(1));
440	&mov	($D,&wparam(2));
441	&mov	(@T[0],"esp");
442
443	# stack frame layout
444	#
445	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
446	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
447	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
448	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
449	#
450	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
451	#	X[4]	X[5]	X[6]	X[7]
452	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
453	#
454	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
455	#	K_40_59	K_40_59	K_40_59	K_40_59
456	#	K_60_79	K_60_79	K_60_79	K_60_79
457	#	K_00_19	K_00_19	K_00_19	K_00_19
458	#	pbswap mask
459	#
460	# +192	ctx				# argument block
461	# +196	inp
462	# +200	end
463	# +204	esp
464	&sub	("esp",208);
465	&and	("esp",-64);
466
467	&movdqa	(&QWP(112+0,"esp"),@X[4]);	# copy constants
468	&movdqa	(&QWP(112+16,"esp"),@X[5]);
469	&movdqa	(&QWP(112+32,"esp"),@X[6]);
470	&shl	($D,6);				# len*64
471	&movdqa	(&QWP(112+48,"esp"),@X[3]);
472	&add	($D,$inp);			# end of input
473	&movdqa	(&QWP(112+64,"esp"),@X[2]);
474	&add	($inp,64);
475	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
476	&mov	(&DWP(192+4,"esp"),$inp);
477	&mov	(&DWP(192+8,"esp"),$D);
478	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
479
480	&mov	($A,&DWP(0,$E));		# load context
481	&mov	($B,&DWP(4,$E));
482	&mov	($C,&DWP(8,$E));
483	&mov	($D,&DWP(12,$E));
484	&mov	($E,&DWP(16,$E));
485	&mov	(@T[0],$B);			# magic seed
486
487	&movdqu	(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
488	&movdqu	(@X[-3&7],&QWP(-48,$inp));
489	&movdqu	(@X[-2&7],&QWP(-32,$inp));
490	&movdqu	(@X[-1&7],&QWP(-16,$inp));
491	&pshufb	(@X[-4&7],@X[2]);		# byte swap
492	&pshufb	(@X[-3&7],@X[2]);
493	&pshufb	(@X[-2&7],@X[2]);
494	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
495	&pshufb	(@X[-1&7],@X[2]);
496	&paddd	(@X[-4&7],@X[3]);		# add K_00_19
497	&paddd	(@X[-3&7],@X[3]);
498	&paddd	(@X[-2&7],@X[3]);
499	&movdqa	(&QWP(0,"esp"),@X[-4&7]);	# X[]+K xfer to IALU
500	&psubd	(@X[-4&7],@X[3]);		# restore X[]
501	&movdqa	(&QWP(0+16,"esp"),@X[-3&7]);
502	&psubd	(@X[-3&7],@X[3]);
503	&movdqa	(&QWP(0+32,"esp"),@X[-2&7]);
504	&psubd	(@X[-2&7],@X[3]);
505	&movdqa	(@X[0],@X[-3&7]);
506	&jmp	(&label("loop"));
507
508######################################################################
509# SSE instruction sequence is first broken to groups of indepentent
510# instructions, independent in respect to their inputs and shifter
511# (not all architectures have more than one). Then IALU instructions
512# are "knitted in" between the SSE groups. Distance is maintained for
513# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
514# [which allegedly also implements SSSE3]...
515#
516# Temporary registers usage. X[2] is volatile at the entry and at the
517# end is restored from backtrace ring buffer. X[3] is expected to
518# contain current K_XX_XX constant and is used to caclulate X[-1]+K
519# from previous round, it becomes volatile the moment the value is
520# saved to stack for transfer to IALU. X[4] becomes volatile whenever
521# X[-4] is accumulated and offloaded to backtrace ring buffer, at the
522# end it is loaded with next K_XX_XX [which becomes X[3] in next
523# round]...
524#
525sub Xupdate_ssse3_16_31()		# recall that $Xi starts wtih 4
526{ use integer;
527  my $body = shift;
528  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
529  my ($a,$b,$c,$d,$e);
530
531	 eval(shift(@insns));
532	 eval(shift(@insns));
533	&palignr(@X[0],@X[-4&7],8);	# compose "X[-14]" in "X[0]"
534	&movdqa	(@X[2],@X[-1&7]);
535	 eval(shift(@insns));
536	 eval(shift(@insns));
537
538	  &paddd	(@X[3],@X[-1&7]);
539	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
540	 eval(shift(@insns));
541	 eval(shift(@insns));
542	&psrldq	(@X[2],4);		# "X[-3]", 3 dwords
543	 eval(shift(@insns));
544	 eval(shift(@insns));
545	&pxor	(@X[0],@X[-4&7]);	# "X[0]"^="X[-16]"
546	 eval(shift(@insns));
547	 eval(shift(@insns));
548
549	&pxor	(@X[2],@X[-2&7]);	# "X[-3]"^"X[-8]"
550	 eval(shift(@insns));
551	 eval(shift(@insns));
552	 eval(shift(@insns));
553	 eval(shift(@insns));
554
555	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
556	 eval(shift(@insns));
557	 eval(shift(@insns));
558	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
559	 eval(shift(@insns));
560	 eval(shift(@insns));
561
562	&movdqa	(@X[4],@X[0]);
563	&movdqa	(@X[2],@X[0]);
564	 eval(shift(@insns));
565	 eval(shift(@insns));
566	 eval(shift(@insns));
567	 eval(shift(@insns));
568
569	&pslldq	(@X[4],12);		# "X[0]"<<96, extract one dword
570	&paddd	(@X[0],@X[0]);
571	 eval(shift(@insns));
572	 eval(shift(@insns));
573	 eval(shift(@insns));
574	 eval(shift(@insns));
575
576	&psrld	(@X[2],31);
577	 eval(shift(@insns));
578	 eval(shift(@insns));
579	&movdqa	(@X[3],@X[4]);
580	 eval(shift(@insns));
581	 eval(shift(@insns));
582
583	&psrld	(@X[4],30);
584	&por	(@X[0],@X[2]);		# "X[0]"<<<=1
585	 eval(shift(@insns));
586	 eval(shift(@insns));
587	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
588	 eval(shift(@insns));
589	 eval(shift(@insns));
590
591	&pslld	(@X[3],2);
592	&pxor	(@X[0],@X[4]);
593	 eval(shift(@insns));
594	 eval(shift(@insns));
595	  &movdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
596	 eval(shift(@insns));
597	 eval(shift(@insns));
598
599	&pxor	(@X[0],@X[3]);		# "X[0]"^=("X[0]"<<96)<<<2
600	  &movdqa	(@X[1],@X[-2&7])	if ($Xi<7);
601	 eval(shift(@insns));
602	 eval(shift(@insns));
603
604	 foreach (@insns) { eval; }	# remaining instructions [if any]
605
606  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
607}
608
609sub Xupdate_ssse3_32_79()
610{ use integer;
611  my $body = shift;
612  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 48 instructions
613  my ($a,$b,$c,$d,$e);
614
615	&movdqa	(@X[2],@X[-1&7])	if ($Xi==8);
616	 eval(shift(@insns));		# body_20_39
617	&pxor	(@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
618	&palignr(@X[2],@X[-2&7],8);	# compose "X[-6]"
619	 eval(shift(@insns));
620	 eval(shift(@insns));
621	 eval(shift(@insns));		# rol
622
623	&pxor	(@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
624	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
625	 eval(shift(@insns));
626	 eval(shift(@insns));
627	 if ($Xi%5) {
628	  &movdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
629	 } else {			# ... or load next one
630	  &movdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
631	 }
632	  &paddd	(@X[3],@X[-1&7]);
633	 eval(shift(@insns));		# ror
634	 eval(shift(@insns));
635
636	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-6]"
637	 eval(shift(@insns));		# body_20_39
638	 eval(shift(@insns));
639	 eval(shift(@insns));
640	 eval(shift(@insns));		# rol
641
642	&movdqa	(@X[2],@X[0]);
643	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
644	 eval(shift(@insns));
645	 eval(shift(@insns));
646	 eval(shift(@insns));		# ror
647	 eval(shift(@insns));
648
649	&pslld	(@X[0],2);
650	 eval(shift(@insns));		# body_20_39
651	 eval(shift(@insns));
652	&psrld	(@X[2],30);
653	 eval(shift(@insns));
654	 eval(shift(@insns));		# rol
655	 eval(shift(@insns));
656	 eval(shift(@insns));
657	 eval(shift(@insns));		# ror
658	 eval(shift(@insns));
659
660	&por	(@X[0],@X[2]);		# "X[0]"<<<=2
661	 eval(shift(@insns));		# body_20_39
662	 eval(shift(@insns));
663	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
664	 eval(shift(@insns));
665	 eval(shift(@insns));		# rol
666	 eval(shift(@insns));
667	 eval(shift(@insns));
668	 eval(shift(@insns));		# ror
669	  &movdqa	(@X[3],@X[0])	if ($Xi<19);
670	 eval(shift(@insns));
671
672	 foreach (@insns) { eval; }	# remaining instructions
673
674  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
675}
676
677sub Xuplast_ssse3_80()
678{ use integer;
679  my $body = shift;
680  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
681  my ($a,$b,$c,$d,$e);
682
683	 eval(shift(@insns));
684	  &paddd	(@X[3],@X[-1&7]);
685	 eval(shift(@insns));
686	 eval(shift(@insns));
687	 eval(shift(@insns));
688	 eval(shift(@insns));
689
690	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
691
692	 foreach (@insns) { eval; }		# remaining instructions
693
694	&mov	($inp=@T[1],&DWP(192+4,"esp"));
695	&cmp	($inp,&DWP(192+8,"esp"));
696	&je	(&label("done"));
697
698	&movdqa	(@X[3],&QWP(112+48,"esp"));	# K_00_19
699	&movdqa	(@X[2],&QWP(112+64,"esp"));	# pbswap mask
700	&movdqu	(@X[-4&7],&QWP(0,$inp));	# load input
701	&movdqu	(@X[-3&7],&QWP(16,$inp));
702	&movdqu	(@X[-2&7],&QWP(32,$inp));
703	&movdqu	(@X[-1&7],&QWP(48,$inp));
704	&add	($inp,64);
705	&pshufb	(@X[-4&7],@X[2]);		# byte swap
706	&mov	(&DWP(192+4,"esp"),$inp);
707	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
708
709  $Xi=0;
710}
711
712sub Xloop_ssse3()
713{ use integer;
714  my $body = shift;
715  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
716  my ($a,$b,$c,$d,$e);
717
718	 eval(shift(@insns));
719	 eval(shift(@insns));
720	&pshufb	(@X[($Xi-3)&7],@X[2]);
721	 eval(shift(@insns));
722	 eval(shift(@insns));
723	&paddd	(@X[($Xi-4)&7],@X[3]);
724	 eval(shift(@insns));
725	 eval(shift(@insns));
726	 eval(shift(@insns));
727	 eval(shift(@insns));
728	&movdqa	(&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]);	# X[]+K xfer to IALU
729	 eval(shift(@insns));
730	 eval(shift(@insns));
731	&psubd	(@X[($Xi-4)&7],@X[3]);
732
733	foreach (@insns) { eval; }
734  $Xi++;
735}
736
737sub Xtail_ssse3()
738{ use integer;
739  my $body = shift;
740  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
741  my ($a,$b,$c,$d,$e);
742
743	foreach (@insns) { eval; }
744}
745
746sub body_00_19 () {
747	(
748	'($a,$b,$c,$d,$e)=@V;'.
749	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
750	'&xor	($c,$d);',
751	'&mov	(@T[1],$a);',	# $b in next round
752	'&$_rol	($a,5);',
753	'&and	(@T[0],$c);',	# ($b&($c^$d))
754	'&xor	($c,$d);',	# restore $c
755	'&xor	(@T[0],$d);',
756	'&add	($e,$a);',
757	'&$_ror	($b,$j?7:2);',	# $b>>>2
758	'&add	($e,@T[0]);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
759	);
760}
761
762sub body_20_39 () {
763	(
764	'($a,$b,$c,$d,$e)=@V;'.
765	'&add	($e,&DWP(4*($j++&15),"esp"));',	# X[]+K xfer
766	'&xor	(@T[0],$d);',	# ($b^$d)
767	'&mov	(@T[1],$a);',	# $b in next round
768	'&$_rol	($a,5);',
769	'&xor	(@T[0],$c);',	# ($b^$d^$c)
770	'&add	($e,$a);',
771	'&$_ror	($b,7);',	# $b>>>2
772	'&add	($e,@T[0]);'	.'unshift(@V,pop(@V)); unshift(@T,pop(@T));'
773	);
774}
775
776sub body_40_59 () {
777	(
778	'($a,$b,$c,$d,$e)=@V;'.
779	'&mov	(@T[1],$c);',
780	'&xor	($c,$d);',
781	'&add	($e,&DWP(4*($j++&15),"esp"));',	# X[]+K xfer
782	'&and	(@T[1],$d);',
783	'&and	(@T[0],$c);',	# ($b&($c^$d))
784	'&$_ror	($b,7);',	# $b>>>2
785	'&add	($e,@T[1]);',
786	'&mov	(@T[1],$a);',	# $b in next round
787	'&$_rol	($a,5);',
788	'&add	($e,@T[0]);',
789	'&xor	($c,$d);',	# restore $c
790	'&add	($e,$a);'	.'unshift(@V,pop(@V)); unshift(@T,pop(@T));'
791	);
792}
793
794&set_label("loop",16);
795	&Xupdate_ssse3_16_31(\&body_00_19);
796	&Xupdate_ssse3_16_31(\&body_00_19);
797	&Xupdate_ssse3_16_31(\&body_00_19);
798	&Xupdate_ssse3_16_31(\&body_00_19);
799	&Xupdate_ssse3_32_79(\&body_00_19);
800	&Xupdate_ssse3_32_79(\&body_20_39);
801	&Xupdate_ssse3_32_79(\&body_20_39);
802	&Xupdate_ssse3_32_79(\&body_20_39);
803	&Xupdate_ssse3_32_79(\&body_20_39);
804	&Xupdate_ssse3_32_79(\&body_20_39);
805	&Xupdate_ssse3_32_79(\&body_40_59);
806	&Xupdate_ssse3_32_79(\&body_40_59);
807	&Xupdate_ssse3_32_79(\&body_40_59);
808	&Xupdate_ssse3_32_79(\&body_40_59);
809	&Xupdate_ssse3_32_79(\&body_40_59);
810	&Xupdate_ssse3_32_79(\&body_20_39);
811	&Xuplast_ssse3_80(\&body_20_39);	# can jump to "done"
812
813				$saved_j=$j; @saved_V=@V;
814
815	&Xloop_ssse3(\&body_20_39);
816	&Xloop_ssse3(\&body_20_39);
817	&Xloop_ssse3(\&body_20_39);
818
819	&mov	(@T[1],&DWP(192,"esp"));	# update context
820	&add	($A,&DWP(0,@T[1]));
821	&add	(@T[0],&DWP(4,@T[1]));		# $b
822	&add	($C,&DWP(8,@T[1]));
823	&mov	(&DWP(0,@T[1]),$A);
824	&add	($D,&DWP(12,@T[1]));
825	&mov	(&DWP(4,@T[1]),@T[0]);
826	&add	($E,&DWP(16,@T[1]));
827	&mov	(&DWP(8,@T[1]),$C);
828	&mov	($B,@T[0]);
829	&mov	(&DWP(12,@T[1]),$D);
830	&mov	(&DWP(16,@T[1]),$E);
831	&movdqa	(@X[0],@X[-3&7]);
832
833	&jmp	(&label("loop"));
834
835&set_label("done",16);		$j=$saved_j; @V=@saved_V;
836
837	&Xtail_ssse3(\&body_20_39);
838	&Xtail_ssse3(\&body_20_39);
839	&Xtail_ssse3(\&body_20_39);
840
841	&mov	(@T[1],&DWP(192,"esp"));	# update context
842	&add	($A,&DWP(0,@T[1]));
843	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
844	&add	(@T[0],&DWP(4,@T[1]));		# $b
845	&add	($C,&DWP(8,@T[1]));
846	&mov	(&DWP(0,@T[1]),$A);
847	&add	($D,&DWP(12,@T[1]));
848	&mov	(&DWP(4,@T[1]),@T[0]);
849	&add	($E,&DWP(16,@T[1]));
850	&mov	(&DWP(8,@T[1]),$C);
851	&mov	(&DWP(12,@T[1]),$D);
852	&mov	(&DWP(16,@T[1]),$E);
853
854&function_end("_sha1_block_data_order_ssse3");
855
856if ($ymm) {
857my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
858my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
859my @V=($A,$B,$C,$D,$E);
860my $j=0;			# hash round
861my @T=($T,$tmp1);
862my $inp;
863
864my $_rol=sub { &shld(@_[0],@_) };
865my $_ror=sub { &shrd(@_[0],@_) };
866
867&function_begin("_sha1_block_data_order_avx");
868	&call	(&label("pic_point"));	# make it PIC!
869	&set_label("pic_point");
870	&blindpop($tmp1);
871	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
872&set_label("avx_shortcut");
873	&vzeroall();
874
875	&vmovdqa(@X[3],&QWP(0,$tmp1));		# K_00_19
876	&vmovdqa(@X[4],&QWP(16,$tmp1));		# K_20_39
877	&vmovdqa(@X[5],&QWP(32,$tmp1));		# K_40_59
878	&vmovdqa(@X[6],&QWP(48,$tmp1));		# K_60_79
879	&vmovdqa(@X[2],&QWP(64,$tmp1));		# pbswap mask
880
881	&mov	($E,&wparam(0));		# load argument block
882	&mov	($inp=@T[1],&wparam(1));
883	&mov	($D,&wparam(2));
884	&mov	(@T[0],"esp");
885
886	# stack frame layout
887	#
888	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
889	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
890	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
891	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
892	#
893	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
894	#	X[4]	X[5]	X[6]	X[7]
895	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
896	#
897	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
898	#	K_40_59	K_40_59	K_40_59	K_40_59
899	#	K_60_79	K_60_79	K_60_79	K_60_79
900	#	K_00_19	K_00_19	K_00_19	K_00_19
901	#	pbswap mask
902	#
903	# +192	ctx				# argument block
904	# +196	inp
905	# +200	end
906	# +204	esp
907	&sub	("esp",208);
908	&and	("esp",-64);
909
910	&vmovdqa(&QWP(112+0,"esp"),@X[4]);	# copy constants
911	&vmovdqa(&QWP(112+16,"esp"),@X[5]);
912	&vmovdqa(&QWP(112+32,"esp"),@X[6]);
913	&shl	($D,6);				# len*64
914	&vmovdqa(&QWP(112+48,"esp"),@X[3]);
915	&add	($D,$inp);			# end of input
916	&vmovdqa(&QWP(112+64,"esp"),@X[2]);
917	&add	($inp,64);
918	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
919	&mov	(&DWP(192+4,"esp"),$inp);
920	&mov	(&DWP(192+8,"esp"),$D);
921	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
922
923	&mov	($A,&DWP(0,$E));		# load context
924	&mov	($B,&DWP(4,$E));
925	&mov	($C,&DWP(8,$E));
926	&mov	($D,&DWP(12,$E));
927	&mov	($E,&DWP(16,$E));
928	&mov	(@T[0],$B);			# magic seed
929
930	&vmovdqu(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
931	&vmovdqu(@X[-3&7],&QWP(-48,$inp));
932	&vmovdqu(@X[-2&7],&QWP(-32,$inp));
933	&vmovdqu(@X[-1&7],&QWP(-16,$inp));
934	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);	# byte swap
935	&vpshufb(@X[-3&7],@X[-3&7],@X[2]);
936	&vpshufb(@X[-2&7],@X[-2&7],@X[2]);
937	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
938	&vpshufb(@X[-1&7],@X[-1&7],@X[2]);
939	&vpaddd	(@X[0],@X[-4&7],@X[3]);		# add K_00_19
940	&vpaddd	(@X[1],@X[-3&7],@X[3]);
941	&vpaddd	(@X[2],@X[-2&7],@X[3]);
942	&vmovdqa(&QWP(0,"esp"),@X[0]);		# X[]+K xfer to IALU
943	&vmovdqa(&QWP(0+16,"esp"),@X[1]);
944	&vmovdqa(&QWP(0+32,"esp"),@X[2]);
945	&jmp	(&label("loop"));
946
947sub Xupdate_avx_16_31()		# recall that $Xi starts wtih 4
948{ use integer;
949  my $body = shift;
950  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
951  my ($a,$b,$c,$d,$e);
952
953	 eval(shift(@insns));
954	 eval(shift(@insns));
955	&vpalignr(@X[0],@X[-3&7],@X[-4&7],8);	# compose "X[-14]" in "X[0]"
956	 eval(shift(@insns));
957	 eval(shift(@insns));
958
959	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
960	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
961	 eval(shift(@insns));
962	 eval(shift(@insns));
963	&vpsrldq(@X[2],@X[-1&7],4);		# "X[-3]", 3 dwords
964	 eval(shift(@insns));
965	 eval(shift(@insns));
966	&vpxor	(@X[0],@X[0],@X[-4&7]);		# "X[0]"^="X[-16]"
967	 eval(shift(@insns));
968	 eval(shift(@insns));
969
970	&vpxor	(@X[2],@X[2],@X[-2&7]);		# "X[-3]"^"X[-8]"
971	 eval(shift(@insns));
972	 eval(shift(@insns));
973	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
974	 eval(shift(@insns));
975	 eval(shift(@insns));
976
977	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
978	 eval(shift(@insns));
979	 eval(shift(@insns));
980	 eval(shift(@insns));
981	 eval(shift(@insns));
982
983	&vpsrld	(@X[2],@X[0],31);
984	 eval(shift(@insns));
985	 eval(shift(@insns));
986	 eval(shift(@insns));
987	 eval(shift(@insns));
988
989	&vpslldq(@X[4],@X[0],12);		# "X[0]"<<96, extract one dword
990	&vpaddd	(@X[0],@X[0],@X[0]);
991	 eval(shift(@insns));
992	 eval(shift(@insns));
993	 eval(shift(@insns));
994	 eval(shift(@insns));
995
996	&vpsrld	(@X[3],@X[4],30);
997	&vpor	(@X[0],@X[0],@X[2]);		# "X[0]"<<<=1
998	 eval(shift(@insns));
999	 eval(shift(@insns));
1000	 eval(shift(@insns));
1001	 eval(shift(@insns));
1002
1003	&vpslld	(@X[4],@X[4],2);
1004	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
1005	 eval(shift(@insns));
1006	 eval(shift(@insns));
1007	&vpxor	(@X[0],@X[0],@X[3]);
1008	 eval(shift(@insns));
1009	 eval(shift(@insns));
1010	 eval(shift(@insns));
1011	 eval(shift(@insns));
1012
1013	&vpxor	(@X[0],@X[0],@X[4]);		# "X[0]"^=("X[0]"<<96)<<<2
1014	 eval(shift(@insns));
1015	 eval(shift(@insns));
1016	  &vmovdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
1017	 eval(shift(@insns));
1018	 eval(shift(@insns));
1019
1020	 foreach (@insns) { eval; }	# remaining instructions [if any]
1021
1022  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1023}
1024
1025sub Xupdate_avx_32_79()
1026{ use integer;
1027  my $body = shift;
1028  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 48 instructions
1029  my ($a,$b,$c,$d,$e);
1030
1031	&vpalignr(@X[2],@X[-1&7],@X[-2&7],8);	# compose "X[-6]"
1032	&vpxor	(@X[0],@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
1033	 eval(shift(@insns));		# body_20_39
1034	 eval(shift(@insns));
1035	 eval(shift(@insns));
1036	 eval(shift(@insns));		# rol
1037
1038	&vpxor	(@X[0],@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
1039	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
1040	 eval(shift(@insns));
1041	 eval(shift(@insns));
1042	 if ($Xi%5) {
1043	  &vmovdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
1044	 } else {			# ... or load next one
1045	  &vmovdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1046	 }
1047	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1048	 eval(shift(@insns));		# ror
1049	 eval(shift(@insns));
1050
1051	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-6]"
1052	 eval(shift(@insns));		# body_20_39
1053	 eval(shift(@insns));
1054	 eval(shift(@insns));
1055	 eval(shift(@insns));		# rol
1056
1057	&vpsrld	(@X[2],@X[0],30);
1058	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1059	 eval(shift(@insns));
1060	 eval(shift(@insns));
1061	 eval(shift(@insns));		# ror
1062	 eval(shift(@insns));
1063
1064	&vpslld	(@X[0],@X[0],2);
1065	 eval(shift(@insns));		# body_20_39
1066	 eval(shift(@insns));
1067	 eval(shift(@insns));
1068	 eval(shift(@insns));		# rol
1069	 eval(shift(@insns));
1070	 eval(shift(@insns));
1071	 eval(shift(@insns));		# ror
1072	 eval(shift(@insns));
1073
1074	&vpor	(@X[0],@X[0],@X[2]);	# "X[0]"<<<=2
1075	 eval(shift(@insns));		# body_20_39
1076	 eval(shift(@insns));
1077	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
1078	 eval(shift(@insns));
1079	 eval(shift(@insns));		# rol
1080	 eval(shift(@insns));
1081	 eval(shift(@insns));
1082	 eval(shift(@insns));		# ror
1083	 eval(shift(@insns));
1084
1085	 foreach (@insns) { eval; }	# remaining instructions
1086
1087  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1088}
1089
1090sub Xuplast_avx_80()
1091{ use integer;
1092  my $body = shift;
1093  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1094  my ($a,$b,$c,$d,$e);
1095
1096	 eval(shift(@insns));
1097	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1098	 eval(shift(@insns));
1099	 eval(shift(@insns));
1100	 eval(shift(@insns));
1101	 eval(shift(@insns));
1102
1103	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
1104
1105	 foreach (@insns) { eval; }		# remaining instructions
1106
1107	&mov	($inp=@T[1],&DWP(192+4,"esp"));
1108	&cmp	($inp,&DWP(192+8,"esp"));
1109	&je	(&label("done"));
1110
1111	&vmovdqa(@X[3],&QWP(112+48,"esp"));	# K_00_19
1112	&vmovdqa(@X[2],&QWP(112+64,"esp"));	# pbswap mask
1113	&vmovdqu(@X[-4&7],&QWP(0,$inp));	# load input
1114	&vmovdqu(@X[-3&7],&QWP(16,$inp));
1115	&vmovdqu(@X[-2&7],&QWP(32,$inp));
1116	&vmovdqu(@X[-1&7],&QWP(48,$inp));
1117	&add	($inp,64);
1118	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);		# byte swap
1119	&mov	(&DWP(192+4,"esp"),$inp);
1120	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
1121
1122  $Xi=0;
1123}
1124
1125sub Xloop_avx()
1126{ use integer;
1127  my $body = shift;
1128  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1129  my ($a,$b,$c,$d,$e);
1130
1131	 eval(shift(@insns));
1132	 eval(shift(@insns));
1133	&vpshufb	(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1134	 eval(shift(@insns));
1135	 eval(shift(@insns));
1136	&vpaddd	(@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1137	 eval(shift(@insns));
1138	 eval(shift(@insns));
1139	 eval(shift(@insns));
1140	 eval(shift(@insns));
1141	&vmovdqa	(&QWP(0+16*$Xi,"esp"),@X[$Xi&7]);	# X[]+K xfer to IALU
1142	 eval(shift(@insns));
1143	 eval(shift(@insns));
1144
1145	foreach (@insns) { eval; }
1146  $Xi++;
1147}
1148
1149sub Xtail_avx()
1150{ use integer;
1151  my $body = shift;
1152  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1153  my ($a,$b,$c,$d,$e);
1154
1155	foreach (@insns) { eval; }
1156}
1157
1158&set_label("loop",16);
1159	&Xupdate_avx_16_31(\&body_00_19);
1160	&Xupdate_avx_16_31(\&body_00_19);
1161	&Xupdate_avx_16_31(\&body_00_19);
1162	&Xupdate_avx_16_31(\&body_00_19);
1163	&Xupdate_avx_32_79(\&body_00_19);
1164	&Xupdate_avx_32_79(\&body_20_39);
1165	&Xupdate_avx_32_79(\&body_20_39);
1166	&Xupdate_avx_32_79(\&body_20_39);
1167	&Xupdate_avx_32_79(\&body_20_39);
1168	&Xupdate_avx_32_79(\&body_20_39);
1169	&Xupdate_avx_32_79(\&body_40_59);
1170	&Xupdate_avx_32_79(\&body_40_59);
1171	&Xupdate_avx_32_79(\&body_40_59);
1172	&Xupdate_avx_32_79(\&body_40_59);
1173	&Xupdate_avx_32_79(\&body_40_59);
1174	&Xupdate_avx_32_79(\&body_20_39);
1175	&Xuplast_avx_80(\&body_20_39);	# can jump to "done"
1176
1177				$saved_j=$j; @saved_V=@V;
1178
1179	&Xloop_avx(\&body_20_39);
1180	&Xloop_avx(\&body_20_39);
1181	&Xloop_avx(\&body_20_39);
1182
1183	&mov	(@T[1],&DWP(192,"esp"));	# update context
1184	&add	($A,&DWP(0,@T[1]));
1185	&add	(@T[0],&DWP(4,@T[1]));		# $b
1186	&add	($C,&DWP(8,@T[1]));
1187	&mov	(&DWP(0,@T[1]),$A);
1188	&add	($D,&DWP(12,@T[1]));
1189	&mov	(&DWP(4,@T[1]),@T[0]);
1190	&add	($E,&DWP(16,@T[1]));
1191	&mov	(&DWP(8,@T[1]),$C);
1192	&mov	($B,@T[0]);
1193	&mov	(&DWP(12,@T[1]),$D);
1194	&mov	(&DWP(16,@T[1]),$E);
1195
1196	&jmp	(&label("loop"));
1197
1198&set_label("done",16);		$j=$saved_j; @V=@saved_V;
1199
1200	&Xtail_avx(\&body_20_39);
1201	&Xtail_avx(\&body_20_39);
1202	&Xtail_avx(\&body_20_39);
1203
1204	&vzeroall();
1205
1206	&mov	(@T[1],&DWP(192,"esp"));	# update context
1207	&add	($A,&DWP(0,@T[1]));
1208	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
1209	&add	(@T[0],&DWP(4,@T[1]));		# $b
1210	&add	($C,&DWP(8,@T[1]));
1211	&mov	(&DWP(0,@T[1]),$A);
1212	&add	($D,&DWP(12,@T[1]));
1213	&mov	(&DWP(4,@T[1]),@T[0]);
1214	&add	($E,&DWP(16,@T[1]));
1215	&mov	(&DWP(8,@T[1]),$C);
1216	&mov	(&DWP(12,@T[1]),$D);
1217	&mov	(&DWP(16,@T[1]),$E);
1218&function_end("_sha1_block_data_order_avx");
1219}
1220&set_label("K_XX_XX",64);
1221&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999);	# K_00_19
1222&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1);	# K_20_39
1223&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc);	# K_40_59
1224&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6);	# K_60_79
1225&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);	# pbswap mask
1226}
1227&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
1228
1229&asm_finish();
1230