1#!/usr/bin/env perl
2
3# ====================================================================
4# [Re]written by Andy Polyakov <appro@openssl.org> 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# March 2014.
83#
84# Add support for Intel SHA Extensions.
85
86######################################################################
87# Current performance is summarized in following table. Numbers are
88# CPU clock cycles spent to process single byte (less is better).
89#
90#		x86		SSSE3		AVX
91# Pentium	15.7		-
92# PIII		11.5		-
93# P4		10.6		-
94# AMD K8	7.1		-
95# Core2		7.3		6.0/+22%	-
96# Westmere	7.3		5.5/+33%	-
97# Sandy Bridge	8.8		6.2/+40%	5.1(**)/+73%
98# Ivy Bridge	7.2		4.8/+51%	4.7(**)/+53%
99# Haswell	6.5		4.3/+51%	4.1(**)/+58%
100# Bulldozer	11.6		6.0/+92%
101# VIA Nano	10.6		7.5/+41%
102# Atom		12.5		9.3(*)/+35%
103# Silvermont	14.5		9.9(*)/+46%
104#
105# (*)	Loop is 1056 instructions long and expected result is ~8.25.
106#	The discrepancy is because of front-end limitations, so
107#	called MS-ROM penalties, and on Silvermont even rotate's
108#	limited parallelism.
109#
110# (**)	As per above comment, the result is for AVX *plus* sh[rl]d.
111
112$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
113push(@INC,"${dir}","${dir}../../perlasm");
114require "x86asm.pl";
115
116&asm_init($ARGV[0],"sha1-586.pl",$ARGV[$#ARGV] eq "386");
117
118$xmm=$ymm=0;
119for (@ARGV) { $xmm=1 if (/-DOPENSSL_IA32_SSE2/); }
120
121# In upstream, this is controlled by shelling out to the compiler to check
122# versions, but BoringSSL is intended to be used with pre-generated perlasm
123# output, so this isn't useful anyway.
124$ymm = 1;
125
126$ymm = 0 unless ($xmm);
127
128$shaext=$xmm;	### set to zero if compiling for 1.0.1
129
130# TODO(davidben): Consider enabling the Intel SHA Extensions code once it's
131# been tested.
132$shaext = 0;
133
134&external_label("OPENSSL_ia32cap_P") if ($xmm);
135
136
137$A="eax";
138$B="ebx";
139$C="ecx";
140$D="edx";
141$E="edi";
142$T="esi";
143$tmp1="ebp";
144
145@V=($A,$B,$C,$D,$E,$T);
146
147$alt=0;	# 1 denotes alternative IALU implementation, which performs
148	# 8% *worse* on P4, same on Westmere and Atom, 2% better on
149	# Sandy Bridge...
150
151sub BODY_00_15
152	{
153	local($n,$a,$b,$c,$d,$e,$f)=@_;
154
155	&comment("00_15 $n");
156
157	&mov($f,$c);			# f to hold F_00_19(b,c,d)
158	 if ($n==0)  { &mov($tmp1,$a); }
159	 else        { &mov($a,$tmp1); }
160	&rotl($tmp1,5);			# tmp1=ROTATE(a,5)
161	 &xor($f,$d);
162	&add($tmp1,$e);			# tmp1+=e;
163	 &mov($e,&swtmp($n%16));	# e becomes volatile and is loaded
164	 				# with xi, also note that e becomes
165					# f in next round...
166	&and($f,$b);
167	&rotr($b,2);			# b=ROTATE(b,30)
168	 &xor($f,$d);			# f holds F_00_19(b,c,d)
169	&lea($tmp1,&DWP(0x5a827999,$tmp1,$e));	# tmp1+=K_00_19+xi
170
171	if ($n==15) { &mov($e,&swtmp(($n+1)%16));# pre-fetch f for next round
172		      &add($f,$tmp1); }	# f+=tmp1
173	else        { &add($tmp1,$f); }	# f becomes a in next round
174	&mov($tmp1,$a)			if ($alt && $n==15);
175	}
176
177sub BODY_16_19
178	{
179	local($n,$a,$b,$c,$d,$e,$f)=@_;
180
181	&comment("16_19 $n");
182
183if ($alt) {
184	&xor($c,$d);
185	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
186	&and($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d), b&=c^d
187	 &xor($f,&swtmp(($n+8)%16));
188	&xor($tmp1,$d);			# tmp1=F_00_19(b,c,d)
189	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
190	&rotl($f,1);			# f=ROTATE(f,1)
191	 &add($e,$tmp1);		# e+=F_00_19(b,c,d)
192	&xor($c,$d);			# restore $c
193	 &mov($tmp1,$a);		# b in next round
194	&rotr($b,$n==16?2:7);		# b=ROTATE(b,30)
195	 &mov(&swtmp($n%16),$f);	# xi=f
196	&rotl($a,5);			# ROTATE(a,5)
197	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
198	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
199	 &add($f,$a);			# f+=ROTATE(a,5)
200} else {
201	&mov($tmp1,$c);			# tmp1 to hold F_00_19(b,c,d)
202	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
203	&xor($tmp1,$d);
204	 &xor($f,&swtmp(($n+8)%16));
205	&and($tmp1,$b);
206	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
207	&rotl($f,1);			# f=ROTATE(f,1)
208	 &xor($tmp1,$d);		# tmp1=F_00_19(b,c,d)
209	&add($e,$tmp1);			# e+=F_00_19(b,c,d)
210	 &mov($tmp1,$a);
211	&rotr($b,2);			# b=ROTATE(b,30)
212	 &mov(&swtmp($n%16),$f);	# xi=f
213	&rotl($tmp1,5);			# ROTATE(a,5)
214	 &lea($f,&DWP(0x5a827999,$f,$e));# f+=F_00_19(b,c,d)+e
215	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
216	 &add($f,$tmp1);		# f+=ROTATE(a,5)
217}
218	}
219
220sub BODY_20_39
221	{
222	local($n,$a,$b,$c,$d,$e,$f)=@_;
223	local $K=($n<40)?0x6ed9eba1:0xca62c1d6;
224
225	&comment("20_39 $n");
226
227if ($alt) {
228	&xor($tmp1,$c);			# tmp1 to hold F_20_39(b,c,d), b^=c
229	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
230	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
231	 &xor($f,&swtmp(($n+8)%16));
232	&add($e,$tmp1);			# e+=F_20_39(b,c,d)
233	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
234	&rotl($f,1);			# f=ROTATE(f,1)
235	 &mov($tmp1,$a);		# b in next round
236	&rotr($b,7);			# b=ROTATE(b,30)
237	 &mov(&swtmp($n%16),$f)		if($n<77);# xi=f
238	&rotl($a,5);			# ROTATE(a,5)
239	 &xor($b,$c)			if($n==39);# warm up for BODY_40_59
240	&and($tmp1,$b)			if($n==39);
241	 &lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
242	&mov($e,&swtmp(($n+1)%16))	if($n<79);# pre-fetch f for next round
243	 &add($f,$a);			# f+=ROTATE(a,5)
244	&rotr($a,5)			if ($n==79);
245} else {
246	&mov($tmp1,$b);			# tmp1 to hold F_20_39(b,c,d)
247	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
248	&xor($tmp1,$c);
249	 &xor($f,&swtmp(($n+8)%16));
250	&xor($tmp1,$d);			# tmp1 holds F_20_39(b,c,d)
251	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
252	&rotl($f,1);			# f=ROTATE(f,1)
253	 &add($e,$tmp1);		# e+=F_20_39(b,c,d)
254	&rotr($b,2);			# b=ROTATE(b,30)
255	 &mov($tmp1,$a);
256	&rotl($tmp1,5);			# ROTATE(a,5)
257	 &mov(&swtmp($n%16),$f) if($n<77);# xi=f
258	&lea($f,&DWP($K,$f,$e));	# f+=e+K_XX_YY
259	 &mov($e,&swtmp(($n+1)%16)) if($n<79);# pre-fetch f for next round
260	&add($f,$tmp1);			# f+=ROTATE(a,5)
261}
262	}
263
264sub BODY_40_59
265	{
266	local($n,$a,$b,$c,$d,$e,$f)=@_;
267
268	&comment("40_59 $n");
269
270if ($alt) {
271	&add($e,$tmp1);			# e+=b&(c^d)
272	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
273	&mov($tmp1,$d);
274	 &xor($f,&swtmp(($n+8)%16));
275	&xor($c,$d);			# restore $c
276	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
277	&rotl($f,1);			# f=ROTATE(f,1)
278	 &and($tmp1,$c);
279	&rotr($b,7);			# b=ROTATE(b,30)
280	 &add($e,$tmp1);		# e+=c&d
281	&mov($tmp1,$a);			# b in next round
282	 &mov(&swtmp($n%16),$f);	# xi=f
283	&rotl($a,5);			# ROTATE(a,5)
284	 &xor($b,$c)			if ($n<59);
285	&and($tmp1,$b)			if ($n<59);# tmp1 to hold F_40_59(b,c,d)
286	 &lea($f,&DWP(0x8f1bbcdc,$f,$e));# f+=K_40_59+e+(b&(c^d))
287	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
288	 &add($f,$a);			# f+=ROTATE(a,5)
289} else {
290	&mov($tmp1,$c);			# tmp1 to hold F_40_59(b,c,d)
291	 &xor($f,&swtmp(($n+2)%16));	# f to hold Xupdate(xi,xa,xb,xc,xd)
292	&xor($tmp1,$d);
293	 &xor($f,&swtmp(($n+8)%16));
294	&and($tmp1,$b);
295	 &xor($f,&swtmp(($n+13)%16));	# f holds xa^xb^xc^xd
296	&rotl($f,1);			# f=ROTATE(f,1)
297	 &add($tmp1,$e);		# b&(c^d)+=e
298	&rotr($b,2);			# b=ROTATE(b,30)
299	 &mov($e,$a);			# e becomes volatile
300	&rotl($e,5);			# ROTATE(a,5)
301	 &mov(&swtmp($n%16),$f);	# xi=f
302	&lea($f,&DWP(0x8f1bbcdc,$f,$tmp1));# f+=K_40_59+e+(b&(c^d))
303	 &mov($tmp1,$c);
304	&add($f,$e);			# f+=ROTATE(a,5)
305	 &and($tmp1,$d);
306	&mov($e,&swtmp(($n+1)%16));	# pre-fetch f for next round
307	 &add($f,$tmp1);		# f+=c&d
308}
309	}
310
311&function_begin("sha1_block_data_order");
312if ($xmm) {
313  &static_label("shaext_shortcut")	if ($shaext);
314  &static_label("ssse3_shortcut");
315  &static_label("avx_shortcut")		if ($ymm);
316  &static_label("K_XX_XX");
317
318	&call	(&label("pic_point"));	# make it PIC!
319  &set_label("pic_point");
320	&blindpop($tmp1);
321	&picmeup($T,"OPENSSL_ia32cap_P",$tmp1,&label("pic_point"));
322	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
323
324	&mov	($A,&DWP(0,$T));
325	&mov	($D,&DWP(4,$T));
326	&test	($D,1<<9);		# check SSSE3 bit
327	&jz	(&label("x86"));
328	&mov	($C,&DWP(8,$T));
329	&test	($A,1<<24);		# check FXSR bit
330	&jz	(&label("x86"));
331	if ($shaext) {
332		&test	($C,1<<29);		# check SHA bit
333		&jnz	(&label("shaext_shortcut"));
334	}
335	if ($ymm) {
336		&and	($D,1<<28);		# mask AVX bit
337		&and	($A,1<<30);		# mask "Intel CPU" bit
338		&or	($A,$D);
339		&cmp	($A,1<<28|1<<30);
340		&je	(&label("avx_shortcut"));
341	}
342	&jmp	(&label("ssse3_shortcut"));
343  &set_label("x86",16);
344}
345	&mov($tmp1,&wparam(0));	# SHA_CTX *c
346	&mov($T,&wparam(1));	# const void *input
347	&mov($A,&wparam(2));	# size_t num
348	&stack_push(16+3);	# allocate X[16]
349	&shl($A,6);
350	&add($A,$T);
351	&mov(&wparam(2),$A);	# pointer beyond the end of input
352	&mov($E,&DWP(16,$tmp1));# pre-load E
353	&jmp(&label("loop"));
354
355&set_label("loop",16);
356
357	# copy input chunk to X, but reversing byte order!
358	for ($i=0; $i<16; $i+=4)
359		{
360		&mov($A,&DWP(4*($i+0),$T));
361		&mov($B,&DWP(4*($i+1),$T));
362		&mov($C,&DWP(4*($i+2),$T));
363		&mov($D,&DWP(4*($i+3),$T));
364		&bswap($A);
365		&bswap($B);
366		&bswap($C);
367		&bswap($D);
368		&mov(&swtmp($i+0),$A);
369		&mov(&swtmp($i+1),$B);
370		&mov(&swtmp($i+2),$C);
371		&mov(&swtmp($i+3),$D);
372		}
373	&mov(&wparam(1),$T);	# redundant in 1st spin
374
375	&mov($A,&DWP(0,$tmp1));	# load SHA_CTX
376	&mov($B,&DWP(4,$tmp1));
377	&mov($C,&DWP(8,$tmp1));
378	&mov($D,&DWP(12,$tmp1));
379	# E is pre-loaded
380
381	for($i=0;$i<16;$i++)	{ &BODY_00_15($i,@V); unshift(@V,pop(@V)); }
382	for(;$i<20;$i++)	{ &BODY_16_19($i,@V); unshift(@V,pop(@V)); }
383	for(;$i<40;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
384	for(;$i<60;$i++)	{ &BODY_40_59($i,@V); unshift(@V,pop(@V)); }
385	for(;$i<80;$i++)	{ &BODY_20_39($i,@V); unshift(@V,pop(@V)); }
386
387	(($V[5] eq $D) and ($V[0] eq $E)) or die;	# double-check
388
389	&mov($tmp1,&wparam(0));	# re-load SHA_CTX*
390	&mov($D,&wparam(1));	# D is last "T" and is discarded
391
392	&add($E,&DWP(0,$tmp1));	# E is last "A"...
393	&add($T,&DWP(4,$tmp1));
394	&add($A,&DWP(8,$tmp1));
395	&add($B,&DWP(12,$tmp1));
396	&add($C,&DWP(16,$tmp1));
397
398	&mov(&DWP(0,$tmp1),$E);	# update SHA_CTX
399	 &add($D,64);		# advance input pointer
400	&mov(&DWP(4,$tmp1),$T);
401	 &cmp($D,&wparam(2));	# have we reached the end yet?
402	&mov(&DWP(8,$tmp1),$A);
403	 &mov($E,$C);		# C is last "E" which needs to be "pre-loaded"
404	&mov(&DWP(12,$tmp1),$B);
405	 &mov($T,$D);		# input pointer
406	&mov(&DWP(16,$tmp1),$C);
407	&jb(&label("loop"));
408
409	&stack_pop(16+3);
410&function_end("sha1_block_data_order");
411
412if ($xmm) {
413if ($shaext) {
414######################################################################
415# Intel SHA Extensions implementation of SHA1 update function.
416#
417my ($ctx,$inp,$num)=("edi","esi","ecx");
418my ($ABCD,$E,$E_,$BSWAP)=map("xmm$_",(0..3));
419my @MSG=map("xmm$_",(4..7));
420
421sub sha1rnds4 {
422 my ($dst,$src,$imm)=@_;
423    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
424    {	&data_byte(0x0f,0x3a,0xcc,0xc0|($1<<3)|$2,$imm);	}
425}
426sub sha1op38 {
427 my ($opcodelet,$dst,$src)=@_;
428    if ("$dst:$src" =~ /xmm([0-7]):xmm([0-7])/)
429    {	&data_byte(0x0f,0x38,$opcodelet,0xc0|($1<<3)|$2);	}
430}
431sub sha1nexte	{ sha1op38(0xc8,@_); }
432sub sha1msg1	{ sha1op38(0xc9,@_); }
433sub sha1msg2	{ sha1op38(0xca,@_); }
434
435&function_begin("_sha1_block_data_order_shaext");
436	&call	(&label("pic_point"));	# make it PIC!
437	&set_label("pic_point");
438	&blindpop($tmp1);
439	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
440&set_label("shaext_shortcut");
441	&mov	($ctx,&wparam(0));
442	&mov	("ebx","esp");
443	&mov	($inp,&wparam(1));
444	&mov	($num,&wparam(2));
445	&sub	("esp",32);
446
447	&movdqu	($ABCD,&QWP(0,$ctx));
448	&movd	($E,&DWP(16,$ctx));
449	&and	("esp",-32);
450	&movdqa	($BSWAP,&QWP(0x50,$tmp1));	# byte-n-word swap
451
452	&movdqu	(@MSG[0],&QWP(0,$inp));
453	&pshufd	($ABCD,$ABCD,0b00011011);	# flip word order
454	&movdqu	(@MSG[1],&QWP(0x10,$inp));
455	&pshufd	($E,$E,0b00011011);		# flip word order
456	&movdqu	(@MSG[2],&QWP(0x20,$inp));
457	&pshufb	(@MSG[0],$BSWAP);
458	&movdqu	(@MSG[3],&QWP(0x30,$inp));
459	&pshufb	(@MSG[1],$BSWAP);
460	&pshufb	(@MSG[2],$BSWAP);
461	&pshufb	(@MSG[3],$BSWAP);
462	&jmp	(&label("loop_shaext"));
463
464&set_label("loop_shaext",16);
465	&dec		($num);
466	&lea		("eax",&DWP(0x40,$inp));
467	&movdqa		(&QWP(0,"esp"),$E);	# offload $E
468	&paddd		($E,@MSG[0]);
469	&cmovne		($inp,"eax");
470	&movdqa		(&QWP(16,"esp"),$ABCD);	# offload $ABCD
471
472for($i=0;$i<20-4;$i+=2) {
473	&sha1msg1	(@MSG[0],@MSG[1]);
474	&movdqa		($E_,$ABCD);
475	&sha1rnds4	($ABCD,$E,int($i/5));	# 0-3...
476	&sha1nexte	($E_,@MSG[1]);
477	&pxor		(@MSG[0],@MSG[2]);
478	&sha1msg1	(@MSG[1],@MSG[2]);
479	&sha1msg2	(@MSG[0],@MSG[3]);
480
481	&movdqa		($E,$ABCD);
482	&sha1rnds4	($ABCD,$E_,int(($i+1)/5));
483	&sha1nexte	($E,@MSG[2]);
484	&pxor		(@MSG[1],@MSG[3]);
485	&sha1msg2	(@MSG[1],@MSG[0]);
486
487	push(@MSG,shift(@MSG));	push(@MSG,shift(@MSG));
488}
489	&movdqu		(@MSG[0],&QWP(0,$inp));
490	&movdqa		($E_,$ABCD);
491	&sha1rnds4	($ABCD,$E,3);		# 64-67
492	&sha1nexte	($E_,@MSG[1]);
493	&movdqu		(@MSG[1],&QWP(0x10,$inp));
494	&pshufb		(@MSG[0],$BSWAP);
495
496	&movdqa		($E,$ABCD);
497	&sha1rnds4	($ABCD,$E_,3);		# 68-71
498	&sha1nexte	($E,@MSG[2]);
499	&movdqu		(@MSG[2],&QWP(0x20,$inp));
500	&pshufb		(@MSG[1],$BSWAP);
501
502	&movdqa		($E_,$ABCD);
503	&sha1rnds4	($ABCD,$E,3);		# 72-75
504	&sha1nexte	($E_,@MSG[3]);
505	&movdqu		(@MSG[3],&QWP(0x30,$inp));
506	&pshufb		(@MSG[2],$BSWAP);
507
508	&movdqa		($E,$ABCD);
509	&sha1rnds4	($ABCD,$E_,3);		# 76-79
510	&movdqa		($E_,&QWP(0,"esp"));
511	&pshufb		(@MSG[3],$BSWAP);
512	&sha1nexte	($E,$E_);
513	&paddd		($ABCD,&QWP(16,"esp"));
514
515	&jnz		(&label("loop_shaext"));
516
517	&pshufd	($ABCD,$ABCD,0b00011011);
518	&pshufd	($E,$E,0b00011011);
519	&movdqu	(&QWP(0,$ctx),$ABCD)
520	&movd	(&DWP(16,$ctx),$E);
521	&mov	("esp","ebx");
522&function_end("_sha1_block_data_order_shaext");
523}
524######################################################################
525# The SSSE3 implementation.
526#
527# %xmm[0-7] are used as ring @X[] buffer containing quadruples of last
528# 32 elements of the message schedule or Xupdate outputs. First 4
529# quadruples are simply byte-swapped input, next 4 are calculated
530# according to method originally suggested by Dean Gaudet (modulo
531# being implemented in SSSE3). Once 8 quadruples or 32 elements are
532# collected, it switches to routine proposed by Max Locktyukhin.
533#
534# Calculations inevitably require temporary reqisters, and there are
535# no %xmm registers left to spare. For this reason part of the ring
536# buffer, X[2..4] to be specific, is offloaded to 3 quadriples ring
537# buffer on the stack. Keep in mind that X[2] is alias X[-6], X[3] -
538# X[-5], and X[4] - X[-4]...
539#
540# Another notable optimization is aggressive stack frame compression
541# aiming to minimize amount of 9-byte instructions...
542#
543# Yet another notable optimization is "jumping" $B variable. It means
544# that there is no register permanently allocated for $B value. This
545# allowed to eliminate one instruction from body_20_39...
546#
547my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
548my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
549my @V=($A,$B,$C,$D,$E);
550my $j=0;			# hash round
551my $rx=0;
552my @T=($T,$tmp1);
553my $inp;
554
555my $_rol=sub { &rol(@_) };
556my $_ror=sub { &ror(@_) };
557
558&function_begin("_sha1_block_data_order_ssse3");
559	&call	(&label("pic_point"));	# make it PIC!
560	&set_label("pic_point");
561	&blindpop($tmp1);
562	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
563&set_label("ssse3_shortcut");
564
565	&movdqa	(@X[3],&QWP(0,$tmp1));		# K_00_19
566	&movdqa	(@X[4],&QWP(16,$tmp1));		# K_20_39
567	&movdqa	(@X[5],&QWP(32,$tmp1));		# K_40_59
568	&movdqa	(@X[6],&QWP(48,$tmp1));		# K_60_79
569	&movdqa	(@X[2],&QWP(64,$tmp1));		# pbswap mask
570
571	&mov	($E,&wparam(0));		# load argument block
572	&mov	($inp=@T[1],&wparam(1));
573	&mov	($D,&wparam(2));
574	&mov	(@T[0],"esp");
575
576	# stack frame layout
577	#
578	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
579	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
580	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
581	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
582	#
583	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
584	#	X[4]	X[5]	X[6]	X[7]
585	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
586	#
587	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
588	#	K_40_59	K_40_59	K_40_59	K_40_59
589	#	K_60_79	K_60_79	K_60_79	K_60_79
590	#	K_00_19	K_00_19	K_00_19	K_00_19
591	#	pbswap mask
592	#
593	# +192	ctx				# argument block
594	# +196	inp
595	# +200	end
596	# +204	esp
597	&sub	("esp",208);
598	&and	("esp",-64);
599
600	&movdqa	(&QWP(112+0,"esp"),@X[4]);	# copy constants
601	&movdqa	(&QWP(112+16,"esp"),@X[5]);
602	&movdqa	(&QWP(112+32,"esp"),@X[6]);
603	&shl	($D,6);				# len*64
604	&movdqa	(&QWP(112+48,"esp"),@X[3]);
605	&add	($D,$inp);			# end of input
606	&movdqa	(&QWP(112+64,"esp"),@X[2]);
607	&add	($inp,64);
608	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
609	&mov	(&DWP(192+4,"esp"),$inp);
610	&mov	(&DWP(192+8,"esp"),$D);
611	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
612
613	&mov	($A,&DWP(0,$E));		# load context
614	&mov	($B,&DWP(4,$E));
615	&mov	($C,&DWP(8,$E));
616	&mov	($D,&DWP(12,$E));
617	&mov	($E,&DWP(16,$E));
618	&mov	(@T[0],$B);			# magic seed
619
620	&movdqu	(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
621	&movdqu	(@X[-3&7],&QWP(-48,$inp));
622	&movdqu	(@X[-2&7],&QWP(-32,$inp));
623	&movdqu	(@X[-1&7],&QWP(-16,$inp));
624	&pshufb	(@X[-4&7],@X[2]);		# byte swap
625	&pshufb	(@X[-3&7],@X[2]);
626	&pshufb	(@X[-2&7],@X[2]);
627	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
628	&pshufb	(@X[-1&7],@X[2]);
629	&paddd	(@X[-4&7],@X[3]);		# add K_00_19
630	&paddd	(@X[-3&7],@X[3]);
631	&paddd	(@X[-2&7],@X[3]);
632	&movdqa	(&QWP(0,"esp"),@X[-4&7]);	# X[]+K xfer to IALU
633	&psubd	(@X[-4&7],@X[3]);		# restore X[]
634	&movdqa	(&QWP(0+16,"esp"),@X[-3&7]);
635	&psubd	(@X[-3&7],@X[3]);
636	&movdqa	(&QWP(0+32,"esp"),@X[-2&7]);
637	&mov	(@T[1],$C);
638	&psubd	(@X[-2&7],@X[3]);
639	&xor	(@T[1],$D);
640	&pshufd	(@X[0],@X[-4&7],0xee);		# was &movdqa	(@X[0],@X[-3&7]);
641	&and	(@T[0],@T[1]);
642	&jmp	(&label("loop"));
643
644######################################################################
645# SSE instruction sequence is first broken to groups of indepentent
646# instructions, independent in respect to their inputs and shifter
647# (not all architectures have more than one). Then IALU instructions
648# are "knitted in" between the SSE groups. Distance is maintained for
649# SSE latency of 2 in hope that it fits better upcoming AMD Bulldozer
650# [which allegedly also implements SSSE3]...
651#
652# Temporary registers usage. X[2] is volatile at the entry and at the
653# end is restored from backtrace ring buffer. X[3] is expected to
654# contain current K_XX_XX constant and is used to caclulate X[-1]+K
655# from previous round, it becomes volatile the moment the value is
656# saved to stack for transfer to IALU. X[4] becomes volatile whenever
657# X[-4] is accumulated and offloaded to backtrace ring buffer, at the
658# end it is loaded with next K_XX_XX [which becomes X[3] in next
659# round]...
660#
661sub Xupdate_ssse3_16_31()		# recall that $Xi starts wtih 4
662{ use integer;
663  my $body = shift;
664  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
665  my ($a,$b,$c,$d,$e);
666
667	 eval(shift(@insns));		# ror
668	 eval(shift(@insns));
669	 eval(shift(@insns));
670	&punpcklqdq(@X[0],@X[-3&7]);	# compose "X[-14]" in "X[0]", was &palignr(@X[0],@X[-4&7],8);
671	&movdqa	(@X[2],@X[-1&7]);
672	 eval(shift(@insns));
673	 eval(shift(@insns));
674
675	  &paddd	(@X[3],@X[-1&7]);
676	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
677	 eval(shift(@insns));		# rol
678	 eval(shift(@insns));
679	&psrldq	(@X[2],4);		# "X[-3]", 3 dwords
680	 eval(shift(@insns));
681	 eval(shift(@insns));
682	&pxor	(@X[0],@X[-4&7]);	# "X[0]"^="X[-16]"
683	 eval(shift(@insns));
684	 eval(shift(@insns));		# ror
685
686	&pxor	(@X[2],@X[-2&7]);	# "X[-3]"^"X[-8]"
687	 eval(shift(@insns));
688	 eval(shift(@insns));
689	 eval(shift(@insns));
690
691	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
692	 eval(shift(@insns));
693	 eval(shift(@insns));		# rol
694	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
695	 eval(shift(@insns));
696	 eval(shift(@insns));
697
698	&movdqa	(@X[4],@X[0]);
699	 eval(shift(@insns));
700	 eval(shift(@insns));
701	 eval(shift(@insns));		# ror
702	&movdqa (@X[2],@X[0]);
703	 eval(shift(@insns));
704
705	&pslldq	(@X[4],12);		# "X[0]"<<96, extract one dword
706	&paddd	(@X[0],@X[0]);
707	 eval(shift(@insns));
708	 eval(shift(@insns));
709
710	&psrld	(@X[2],31);
711	 eval(shift(@insns));
712	 eval(shift(@insns));		# rol
713	&movdqa	(@X[3],@X[4]);
714	 eval(shift(@insns));
715	 eval(shift(@insns));
716	 eval(shift(@insns));
717
718	&psrld	(@X[4],30);
719	 eval(shift(@insns));
720	 eval(shift(@insns));		# ror
721	&por	(@X[0],@X[2]);		# "X[0]"<<<=1
722	 eval(shift(@insns));
723	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
724	 eval(shift(@insns));
725	 eval(shift(@insns));
726
727	&pslld	(@X[3],2);
728	 eval(shift(@insns));
729	 eval(shift(@insns));		# rol
730	&pxor   (@X[0],@X[4]);
731	  &movdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
732	 eval(shift(@insns));
733	 eval(shift(@insns));
734
735	&pxor	(@X[0],@X[3]);		# "X[0]"^=("X[0]"<<96)<<<2
736	  &pshufd	(@X[1],@X[-3&7],0xee)	if ($Xi<7);	# was &movdqa	(@X[1],@X[-2&7])
737	  &pshufd	(@X[3],@X[-1&7],0xee)	if ($Xi==7);
738	 eval(shift(@insns));
739	 eval(shift(@insns));
740
741	 foreach (@insns) { eval; }	# remaining instructions [if any]
742
743  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
744}
745
746sub Xupdate_ssse3_32_79()
747{ use integer;
748  my $body = shift;
749  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 44 instructions
750  my ($a,$b,$c,$d,$e);
751
752	 eval(shift(@insns));		# body_20_39
753	&pxor	(@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
754	&punpcklqdq(@X[2],@X[-1&7]);	# compose "X[-6]", was &palignr(@X[2],@X[-2&7],8)
755	 eval(shift(@insns));
756	 eval(shift(@insns));
757	 eval(shift(@insns));		# rol
758
759	&pxor	(@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
760	  &movdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
761	 eval(shift(@insns));
762	 eval(shift(@insns));
763	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
764	 if ($Xi%5) {
765	  &movdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
766	 } else {			# ... or load next one
767	  &movdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
768	 }
769	 eval(shift(@insns));		# ror
770	  &paddd	(@X[3],@X[-1&7]);
771	 eval(shift(@insns));
772
773	&pxor	(@X[0],@X[2]);		# "X[0]"^="X[-6]"
774	 eval(shift(@insns));		# body_20_39
775	 eval(shift(@insns));
776	 eval(shift(@insns));
777	 eval(shift(@insns));		# rol
778
779	&movdqa	(@X[2],@X[0]);
780	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
781	 eval(shift(@insns));
782	 eval(shift(@insns));
783	 eval(shift(@insns));		# ror
784	 eval(shift(@insns));
785	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
786
787	&pslld	(@X[0],2);
788	 eval(shift(@insns));		# body_20_39
789	 eval(shift(@insns));
790	&psrld	(@X[2],30);
791	 eval(shift(@insns));
792	 eval(shift(@insns));		# rol
793	 eval(shift(@insns));
794	 eval(shift(@insns));
795	 eval(shift(@insns));		# ror
796	 eval(shift(@insns));
797	 eval(shift(@insns))		if (@insns[1] =~ /_rol/);
798	 eval(shift(@insns))		if (@insns[0] =~ /_rol/);
799
800	&por	(@X[0],@X[2]);		# "X[0]"<<<=2
801	 eval(shift(@insns));		# body_20_39
802	 eval(shift(@insns));
803	  &movdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
804	 eval(shift(@insns));
805	 eval(shift(@insns));		# rol
806	 eval(shift(@insns));
807	 eval(shift(@insns));
808	 eval(shift(@insns));		# ror
809	  &pshufd	(@X[3],@X[-1],0xee)	if ($Xi<19);	# was &movdqa	(@X[3],@X[0])
810	 eval(shift(@insns));
811
812	 foreach (@insns) { eval; }	# remaining instructions
813
814  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
815}
816
817sub Xuplast_ssse3_80()
818{ use integer;
819  my $body = shift;
820  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
821  my ($a,$b,$c,$d,$e);
822
823	 eval(shift(@insns));
824	 eval(shift(@insns));
825	 eval(shift(@insns));
826	 eval(shift(@insns));
827	 eval(shift(@insns));
828	 eval(shift(@insns));
829	 eval(shift(@insns));
830	  &paddd	(@X[3],@X[-1&7]);
831	 eval(shift(@insns));
832	 eval(shift(@insns));
833	 eval(shift(@insns));
834	 eval(shift(@insns));
835
836	  &movdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
837
838	 foreach (@insns) { eval; }		# remaining instructions
839
840	&mov	($inp=@T[1],&DWP(192+4,"esp"));
841	&cmp	($inp,&DWP(192+8,"esp"));
842	&je	(&label("done"));
843
844	&movdqa	(@X[3],&QWP(112+48,"esp"));	# K_00_19
845	&movdqa	(@X[2],&QWP(112+64,"esp"));	# pbswap mask
846	&movdqu	(@X[-4&7],&QWP(0,$inp));	# load input
847	&movdqu	(@X[-3&7],&QWP(16,$inp));
848	&movdqu	(@X[-2&7],&QWP(32,$inp));
849	&movdqu	(@X[-1&7],&QWP(48,$inp));
850	&add	($inp,64);
851	&pshufb	(@X[-4&7],@X[2]);		# byte swap
852	&mov	(&DWP(192+4,"esp"),$inp);
853	&movdqa	(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
854
855  $Xi=0;
856}
857
858sub Xloop_ssse3()
859{ use integer;
860  my $body = shift;
861  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
862  my ($a,$b,$c,$d,$e);
863
864	 eval(shift(@insns));
865	 eval(shift(@insns));
866	 eval(shift(@insns));
867	 eval(shift(@insns));
868	 eval(shift(@insns));
869	 eval(shift(@insns));
870	 eval(shift(@insns));
871	&pshufb	(@X[($Xi-3)&7],@X[2]);
872	 eval(shift(@insns));
873	 eval(shift(@insns));
874	 eval(shift(@insns));
875	 eval(shift(@insns));
876	&paddd	(@X[($Xi-4)&7],@X[3]);
877	 eval(shift(@insns));
878	 eval(shift(@insns));
879	 eval(shift(@insns));
880	 eval(shift(@insns));
881	&movdqa	(&QWP(0+16*$Xi,"esp"),@X[($Xi-4)&7]);	# X[]+K xfer to IALU
882	 eval(shift(@insns));
883	 eval(shift(@insns));
884	 eval(shift(@insns));
885	 eval(shift(@insns));
886	&psubd	(@X[($Xi-4)&7],@X[3]);
887
888	foreach (@insns) { eval; }
889  $Xi++;
890}
891
892sub Xtail_ssse3()
893{ use integer;
894  my $body = shift;
895  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
896  my ($a,$b,$c,$d,$e);
897
898	foreach (@insns) { eval; }
899}
900
901sub body_00_19 () {	# ((c^d)&b)^d
902	# on start @T[0]=(c^d)&b
903	return &body_20_39()	if ($rx==19);	$rx++;
904	(
905	'($a,$b,$c,$d,$e)=@V;'.
906	'&$_ror	($b,$j?7:2);',	# $b>>>2
907	'&xor	(@T[0],$d);',
908	'&mov	(@T[1],$a);',	# $b in next round
909
910	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
911	'&xor	($b,$c);',	# $c^$d for next round
912
913	'&$_rol	($a,5);',
914	'&add	($e,@T[0]);',
915	'&and	(@T[1],$b);',	# ($b&($c^$d)) for next round
916
917	'&xor	($b,$c);',	# restore $b
918	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
919	);
920}
921
922sub body_20_39 () {	# b^d^c
923	# on entry @T[0]=b^d
924	return &body_40_59()	if ($rx==39);	$rx++;
925	(
926	'($a,$b,$c,$d,$e)=@V;'.
927	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
928	'&xor	(@T[0],$d)	if($j==19);'.
929	'&xor	(@T[0],$c)	if($j> 19);',	# ($b^$d^$c)
930	'&mov	(@T[1],$a);',	# $b in next round
931
932	'&$_rol	($a,5);',
933	'&add	($e,@T[0]);',
934	'&xor	(@T[1],$c)	if ($j< 79);',	# $b^$d for next round
935
936	'&$_ror	($b,7);',	# $b>>>2
937	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
938	);
939}
940
941sub body_40_59 () {	# ((b^c)&(c^d))^c
942	# on entry @T[0]=(b^c), (c^=d)
943	$rx++;
944	(
945	'($a,$b,$c,$d,$e)=@V;'.
946	'&add	($e,&DWP(4*($j&15),"esp"));',	# X[]+K xfer
947	'&and	(@T[0],$c)	if ($j>=40);',	# (b^c)&(c^d)
948	'&xor	($c,$d)		if ($j>=40);',	# restore $c
949
950	'&$_ror	($b,7);',	# $b>>>2
951	'&mov	(@T[1],$a);',	# $b for next round
952	'&xor	(@T[0],$c);',
953
954	'&$_rol	($a,5);',
955	'&add	($e,@T[0]);',
956	'&xor	(@T[1],$c)	if ($j==59);'.
957	'&xor	(@T[1],$b)	if ($j< 59);',	# b^c for next round
958
959	'&xor	($b,$c)		if ($j< 59);',	# c^d for next round
960	'&add	($e,$a);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
961	);
962}
963######
964sub bodyx_00_19 () {	# ((c^d)&b)^d
965	# on start @T[0]=(b&c)^(~b&d), $e+=X[]+K
966	return &bodyx_20_39()	if ($rx==19);	$rx++;
967	(
968	'($a,$b,$c,$d,$e)=@V;'.
969
970	'&rorx	($b,$b,2)			if ($j==0);'.	# $b>>>2
971	'&rorx	($b,@T[1],7)			if ($j!=0);',	# $b>>>2
972	'&lea	($e,&DWP(0,$e,@T[0]));',
973	'&rorx	(@T[0],$a,5);',
974
975	'&andn	(@T[1],$a,$c);',
976	'&and	($a,$b)',
977	'&add	($d,&DWP(4*(($j+1)&15),"esp"));',	# X[]+K xfer
978
979	'&xor	(@T[1],$a)',
980	'&add	($e,@T[0]);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
981	);
982}
983
984sub bodyx_20_39 () {	# b^d^c
985	# on start $b=b^c^d
986	return &bodyx_40_59()	if ($rx==39);	$rx++;
987	(
988	'($a,$b,$c,$d,$e)=@V;'.
989
990	'&add	($e,($j==19?@T[0]:$b))',
991	'&rorx	($b,@T[1],7);',	# $b>>>2
992	'&rorx	(@T[0],$a,5);',
993
994	'&xor	($a,$b)				if ($j<79);',
995	'&add	($d,&DWP(4*(($j+1)&15),"esp"))	if ($j<79);',	# X[]+K xfer
996	'&xor	($a,$c)				if ($j<79);',
997	'&add	($e,@T[0]);'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
998	);
999}
1000
1001sub bodyx_40_59 () {	# ((b^c)&(c^d))^c
1002	# on start $b=((b^c)&(c^d))^c
1003	return &bodyx_20_39()	if ($rx==59);	$rx++;
1004	(
1005	'($a,$b,$c,$d,$e)=@V;'.
1006
1007	'&rorx	(@T[0],$a,5)',
1008	'&lea	($e,&DWP(0,$e,$b))',
1009	'&rorx	($b,@T[1],7)',	# $b>>>2
1010	'&add	($d,&DWP(4*(($j+1)&15),"esp"))',	# X[]+K xfer
1011
1012	'&mov	(@T[1],$c)',
1013	'&xor	($a,$b)',	# b^c for next round
1014	'&xor	(@T[1],$b)',	# c^d for next round
1015
1016	'&and	($a,@T[1])',
1017	'&add	($e,@T[0])',
1018	'&xor	($a,$b)'	.'$j++; unshift(@V,pop(@V)); unshift(@T,pop(@T));'
1019	);
1020}
1021
1022&set_label("loop",16);
1023	&Xupdate_ssse3_16_31(\&body_00_19);
1024	&Xupdate_ssse3_16_31(\&body_00_19);
1025	&Xupdate_ssse3_16_31(\&body_00_19);
1026	&Xupdate_ssse3_16_31(\&body_00_19);
1027	&Xupdate_ssse3_32_79(\&body_00_19);
1028	&Xupdate_ssse3_32_79(\&body_20_39);
1029	&Xupdate_ssse3_32_79(\&body_20_39);
1030	&Xupdate_ssse3_32_79(\&body_20_39);
1031	&Xupdate_ssse3_32_79(\&body_20_39);
1032	&Xupdate_ssse3_32_79(\&body_20_39);
1033	&Xupdate_ssse3_32_79(\&body_40_59);
1034	&Xupdate_ssse3_32_79(\&body_40_59);
1035	&Xupdate_ssse3_32_79(\&body_40_59);
1036	&Xupdate_ssse3_32_79(\&body_40_59);
1037	&Xupdate_ssse3_32_79(\&body_40_59);
1038	&Xupdate_ssse3_32_79(\&body_20_39);
1039	&Xuplast_ssse3_80(\&body_20_39);	# can jump to "done"
1040
1041				$saved_j=$j; @saved_V=@V;
1042
1043	&Xloop_ssse3(\&body_20_39);
1044	&Xloop_ssse3(\&body_20_39);
1045	&Xloop_ssse3(\&body_20_39);
1046
1047	&mov	(@T[1],&DWP(192,"esp"));	# update context
1048	&add	($A,&DWP(0,@T[1]));
1049	&add	(@T[0],&DWP(4,@T[1]));		# $b
1050	&add	($C,&DWP(8,@T[1]));
1051	&mov	(&DWP(0,@T[1]),$A);
1052	&add	($D,&DWP(12,@T[1]));
1053	&mov	(&DWP(4,@T[1]),@T[0]);
1054	&add	($E,&DWP(16,@T[1]));
1055	&mov	(&DWP(8,@T[1]),$C);
1056	&mov	($B,$C);
1057	&mov	(&DWP(12,@T[1]),$D);
1058	&xor	($B,$D);
1059	&mov	(&DWP(16,@T[1]),$E);
1060	&mov	(@T[1],@T[0]);
1061	&pshufd	(@X[0],@X[-4&7],0xee);		# was &movdqa	(@X[0],@X[-3&7]);
1062	&and	(@T[0],$B);
1063	&mov	($B,$T[1]);
1064
1065	&jmp	(&label("loop"));
1066
1067&set_label("done",16);		$j=$saved_j; @V=@saved_V;
1068
1069	&Xtail_ssse3(\&body_20_39);
1070	&Xtail_ssse3(\&body_20_39);
1071	&Xtail_ssse3(\&body_20_39);
1072
1073	&mov	(@T[1],&DWP(192,"esp"));	# update context
1074	&add	($A,&DWP(0,@T[1]));
1075	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
1076	&add	(@T[0],&DWP(4,@T[1]));		# $b
1077	&add	($C,&DWP(8,@T[1]));
1078	&mov	(&DWP(0,@T[1]),$A);
1079	&add	($D,&DWP(12,@T[1]));
1080	&mov	(&DWP(4,@T[1]),@T[0]);
1081	&add	($E,&DWP(16,@T[1]));
1082	&mov	(&DWP(8,@T[1]),$C);
1083	&mov	(&DWP(12,@T[1]),$D);
1084	&mov	(&DWP(16,@T[1]),$E);
1085
1086&function_end("_sha1_block_data_order_ssse3");
1087
1088$rx=0;	# reset
1089
1090if ($ymm) {
1091my $Xi=4;			# 4xSIMD Xupdate round, start pre-seeded
1092my @X=map("xmm$_",(4..7,0..3));	# pre-seeded for $Xi=4
1093my @V=($A,$B,$C,$D,$E);
1094my $j=0;			# hash round
1095my @T=($T,$tmp1);
1096my $inp;
1097
1098my $_rol=sub { &shld(@_[0],@_) };
1099my $_ror=sub { &shrd(@_[0],@_) };
1100
1101&function_begin("_sha1_block_data_order_avx");
1102	&call	(&label("pic_point"));	# make it PIC!
1103	&set_label("pic_point");
1104	&blindpop($tmp1);
1105	&lea	($tmp1,&DWP(&label("K_XX_XX")."-".&label("pic_point"),$tmp1));
1106&set_label("avx_shortcut");
1107	&vzeroall();
1108
1109	&vmovdqa(@X[3],&QWP(0,$tmp1));		# K_00_19
1110	&vmovdqa(@X[4],&QWP(16,$tmp1));		# K_20_39
1111	&vmovdqa(@X[5],&QWP(32,$tmp1));		# K_40_59
1112	&vmovdqa(@X[6],&QWP(48,$tmp1));		# K_60_79
1113	&vmovdqa(@X[2],&QWP(64,$tmp1));		# pbswap mask
1114
1115	&mov	($E,&wparam(0));		# load argument block
1116	&mov	($inp=@T[1],&wparam(1));
1117	&mov	($D,&wparam(2));
1118	&mov	(@T[0],"esp");
1119
1120	# stack frame layout
1121	#
1122	# +0	X[0]+K	X[1]+K	X[2]+K	X[3]+K	# XMM->IALU xfer area
1123	#	X[4]+K	X[5]+K	X[6]+K	X[7]+K
1124	#	X[8]+K	X[9]+K	X[10]+K	X[11]+K
1125	#	X[12]+K	X[13]+K	X[14]+K	X[15]+K
1126	#
1127	# +64	X[0]	X[1]	X[2]	X[3]	# XMM->XMM backtrace area
1128	#	X[4]	X[5]	X[6]	X[7]
1129	#	X[8]	X[9]	X[10]	X[11]	# even borrowed for K_00_19
1130	#
1131	# +112	K_20_39	K_20_39	K_20_39	K_20_39	# constants
1132	#	K_40_59	K_40_59	K_40_59	K_40_59
1133	#	K_60_79	K_60_79	K_60_79	K_60_79
1134	#	K_00_19	K_00_19	K_00_19	K_00_19
1135	#	pbswap mask
1136	#
1137	# +192	ctx				# argument block
1138	# +196	inp
1139	# +200	end
1140	# +204	esp
1141	&sub	("esp",208);
1142	&and	("esp",-64);
1143
1144	&vmovdqa(&QWP(112+0,"esp"),@X[4]);	# copy constants
1145	&vmovdqa(&QWP(112+16,"esp"),@X[5]);
1146	&vmovdqa(&QWP(112+32,"esp"),@X[6]);
1147	&shl	($D,6);				# len*64
1148	&vmovdqa(&QWP(112+48,"esp"),@X[3]);
1149	&add	($D,$inp);			# end of input
1150	&vmovdqa(&QWP(112+64,"esp"),@X[2]);
1151	&add	($inp,64);
1152	&mov	(&DWP(192+0,"esp"),$E);		# save argument block
1153	&mov	(&DWP(192+4,"esp"),$inp);
1154	&mov	(&DWP(192+8,"esp"),$D);
1155	&mov	(&DWP(192+12,"esp"),@T[0]);	# save original %esp
1156
1157	&mov	($A,&DWP(0,$E));		# load context
1158	&mov	($B,&DWP(4,$E));
1159	&mov	($C,&DWP(8,$E));
1160	&mov	($D,&DWP(12,$E));
1161	&mov	($E,&DWP(16,$E));
1162	&mov	(@T[0],$B);			# magic seed
1163
1164	&vmovdqu(@X[-4&7],&QWP(-64,$inp));	# load input to %xmm[0-3]
1165	&vmovdqu(@X[-3&7],&QWP(-48,$inp));
1166	&vmovdqu(@X[-2&7],&QWP(-32,$inp));
1167	&vmovdqu(@X[-1&7],&QWP(-16,$inp));
1168	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);	# byte swap
1169	&vpshufb(@X[-3&7],@X[-3&7],@X[2]);
1170	&vpshufb(@X[-2&7],@X[-2&7],@X[2]);
1171	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
1172	&vpshufb(@X[-1&7],@X[-1&7],@X[2]);
1173	&vpaddd	(@X[0],@X[-4&7],@X[3]);		# add K_00_19
1174	&vpaddd	(@X[1],@X[-3&7],@X[3]);
1175	&vpaddd	(@X[2],@X[-2&7],@X[3]);
1176	&vmovdqa(&QWP(0,"esp"),@X[0]);		# X[]+K xfer to IALU
1177	&mov	(@T[1],$C);
1178	&vmovdqa(&QWP(0+16,"esp"),@X[1]);
1179	&xor	(@T[1],$D);
1180	&vmovdqa(&QWP(0+32,"esp"),@X[2]);
1181	&and	(@T[0],@T[1]);
1182	&jmp	(&label("loop"));
1183
1184sub Xupdate_avx_16_31()		# recall that $Xi starts wtih 4
1185{ use integer;
1186  my $body = shift;
1187  my @insns = (&$body,&$body,&$body,&$body);	# 40 instructions
1188  my ($a,$b,$c,$d,$e);
1189
1190	 eval(shift(@insns));
1191	 eval(shift(@insns));
1192	&vpalignr(@X[0],@X[-3&7],@X[-4&7],8);	# compose "X[-14]" in "X[0]"
1193	 eval(shift(@insns));
1194	 eval(shift(@insns));
1195
1196	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1197	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);# save X[] to backtrace buffer
1198	 eval(shift(@insns));
1199	 eval(shift(@insns));
1200	&vpsrldq(@X[2],@X[-1&7],4);		# "X[-3]", 3 dwords
1201	 eval(shift(@insns));
1202	 eval(shift(@insns));
1203	&vpxor	(@X[0],@X[0],@X[-4&7]);		# "X[0]"^="X[-16]"
1204	 eval(shift(@insns));
1205	 eval(shift(@insns));
1206
1207	&vpxor	(@X[2],@X[2],@X[-2&7]);		# "X[-3]"^"X[-8]"
1208	 eval(shift(@insns));
1209	 eval(shift(@insns));
1210	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1211	 eval(shift(@insns));
1212	 eval(shift(@insns));
1213
1214	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-3]"^"X[-8]"
1215	 eval(shift(@insns));
1216	 eval(shift(@insns));
1217	 eval(shift(@insns));
1218	 eval(shift(@insns));
1219
1220	&vpsrld	(@X[2],@X[0],31);
1221	 eval(shift(@insns));
1222	 eval(shift(@insns));
1223	 eval(shift(@insns));
1224	 eval(shift(@insns));
1225
1226	&vpslldq(@X[4],@X[0],12);		# "X[0]"<<96, extract one dword
1227	&vpaddd	(@X[0],@X[0],@X[0]);
1228	 eval(shift(@insns));
1229	 eval(shift(@insns));
1230	 eval(shift(@insns));
1231	 eval(shift(@insns));
1232
1233	&vpsrld	(@X[3],@X[4],30);
1234	&vpor	(@X[0],@X[0],@X[2]);		# "X[0]"<<<=1
1235	 eval(shift(@insns));
1236	 eval(shift(@insns));
1237	 eval(shift(@insns));
1238	 eval(shift(@insns));
1239
1240	&vpslld	(@X[4],@X[4],2);
1241	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if ($Xi>5);	# restore X[] from backtrace buffer
1242	 eval(shift(@insns));
1243	 eval(shift(@insns));
1244	&vpxor	(@X[0],@X[0],@X[3]);
1245	 eval(shift(@insns));
1246	 eval(shift(@insns));
1247	 eval(shift(@insns));
1248	 eval(shift(@insns));
1249
1250	&vpxor	(@X[0],@X[0],@X[4]);		# "X[0]"^=("X[0]"<<96)<<<2
1251	 eval(shift(@insns));
1252	 eval(shift(@insns));
1253	  &vmovdqa	(@X[4],&QWP(112-16+16*(($Xi)/5),"esp"));	# K_XX_XX
1254	 eval(shift(@insns));
1255	 eval(shift(@insns));
1256
1257	 foreach (@insns) { eval; }	# remaining instructions [if any]
1258
1259  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1260}
1261
1262sub Xupdate_avx_32_79()
1263{ use integer;
1264  my $body = shift;
1265  my @insns = (&$body,&$body,&$body,&$body);	# 32 to 44 instructions
1266  my ($a,$b,$c,$d,$e);
1267
1268	&vpalignr(@X[2],@X[-1&7],@X[-2&7],8);	# compose "X[-6]"
1269	&vpxor	(@X[0],@X[0],@X[-4&7]);	# "X[0]"="X[-32]"^"X[-16]"
1270	 eval(shift(@insns));		# body_20_39
1271	 eval(shift(@insns));
1272	 eval(shift(@insns));
1273	 eval(shift(@insns));		# rol
1274
1275	&vpxor	(@X[0],@X[0],@X[-7&7]);	# "X[0]"^="X[-28]"
1276	  &vmovdqa	(&QWP(64+16*(($Xi-4)%3),"esp"),@X[-4&7]);	# save X[] to backtrace buffer
1277	 eval(shift(@insns));
1278	 eval(shift(@insns));
1279	 if ($Xi%5) {
1280	  &vmovdqa	(@X[4],@X[3]);	# "perpetuate" K_XX_XX...
1281	 } else {			# ... or load next one
1282	  &vmovdqa	(@X[4],&QWP(112-16+16*($Xi/5),"esp"));
1283	 }
1284	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1285	 eval(shift(@insns));		# ror
1286	 eval(shift(@insns));
1287
1288	&vpxor	(@X[0],@X[0],@X[2]);		# "X[0]"^="X[-6]"
1289	 eval(shift(@insns));		# body_20_39
1290	 eval(shift(@insns));
1291	 eval(shift(@insns));
1292	 eval(shift(@insns));		# rol
1293
1294	&vpsrld	(@X[2],@X[0],30);
1295	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer to IALU
1296	 eval(shift(@insns));
1297	 eval(shift(@insns));
1298	 eval(shift(@insns));		# ror
1299	 eval(shift(@insns));
1300
1301	&vpslld	(@X[0],@X[0],2);
1302	 eval(shift(@insns));		# body_20_39
1303	 eval(shift(@insns));
1304	 eval(shift(@insns));
1305	 eval(shift(@insns));		# rol
1306	 eval(shift(@insns));
1307	 eval(shift(@insns));
1308	 eval(shift(@insns));		# ror
1309	 eval(shift(@insns));
1310
1311	&vpor	(@X[0],@X[0],@X[2]);	# "X[0]"<<<=2
1312	 eval(shift(@insns));		# body_20_39
1313	 eval(shift(@insns));
1314	  &vmovdqa	(@X[2],&QWP(64+16*(($Xi-6)%3),"esp")) if($Xi<19);	# restore X[] from backtrace buffer
1315	 eval(shift(@insns));
1316	 eval(shift(@insns));		# rol
1317	 eval(shift(@insns));
1318	 eval(shift(@insns));
1319	 eval(shift(@insns));		# ror
1320	 eval(shift(@insns));
1321
1322	 foreach (@insns) { eval; }	# remaining instructions
1323
1324  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
1325}
1326
1327sub Xuplast_avx_80()
1328{ use integer;
1329  my $body = shift;
1330  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1331  my ($a,$b,$c,$d,$e);
1332
1333	 eval(shift(@insns));
1334	  &vpaddd	(@X[3],@X[3],@X[-1&7]);
1335	 eval(shift(@insns));
1336	 eval(shift(@insns));
1337	 eval(shift(@insns));
1338	 eval(shift(@insns));
1339
1340	  &vmovdqa	(&QWP(0+16*(($Xi-1)&3),"esp"),@X[3]);	# X[]+K xfer IALU
1341
1342	 foreach (@insns) { eval; }		# remaining instructions
1343
1344	&mov	($inp=@T[1],&DWP(192+4,"esp"));
1345	&cmp	($inp,&DWP(192+8,"esp"));
1346	&je	(&label("done"));
1347
1348	&vmovdqa(@X[3],&QWP(112+48,"esp"));	# K_00_19
1349	&vmovdqa(@X[2],&QWP(112+64,"esp"));	# pbswap mask
1350	&vmovdqu(@X[-4&7],&QWP(0,$inp));	# load input
1351	&vmovdqu(@X[-3&7],&QWP(16,$inp));
1352	&vmovdqu(@X[-2&7],&QWP(32,$inp));
1353	&vmovdqu(@X[-1&7],&QWP(48,$inp));
1354	&add	($inp,64);
1355	&vpshufb(@X[-4&7],@X[-4&7],@X[2]);		# byte swap
1356	&mov	(&DWP(192+4,"esp"),$inp);
1357	&vmovdqa(&QWP(112-16,"esp"),@X[3]);	# borrow last backtrace slot
1358
1359  $Xi=0;
1360}
1361
1362sub Xloop_avx()
1363{ use integer;
1364  my $body = shift;
1365  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1366  my ($a,$b,$c,$d,$e);
1367
1368	 eval(shift(@insns));
1369	 eval(shift(@insns));
1370	&vpshufb	(@X[($Xi-3)&7],@X[($Xi-3)&7],@X[2]);
1371	 eval(shift(@insns));
1372	 eval(shift(@insns));
1373	&vpaddd	(@X[$Xi&7],@X[($Xi-4)&7],@X[3]);
1374	 eval(shift(@insns));
1375	 eval(shift(@insns));
1376	 eval(shift(@insns));
1377	 eval(shift(@insns));
1378	&vmovdqa	(&QWP(0+16*$Xi,"esp"),@X[$Xi&7]);	# X[]+K xfer to IALU
1379	 eval(shift(@insns));
1380	 eval(shift(@insns));
1381
1382	foreach (@insns) { eval; }
1383  $Xi++;
1384}
1385
1386sub Xtail_avx()
1387{ use integer;
1388  my $body = shift;
1389  my @insns = (&$body,&$body,&$body,&$body);	# 32 instructions
1390  my ($a,$b,$c,$d,$e);
1391
1392	foreach (@insns) { eval; }
1393}
1394
1395&set_label("loop",16);
1396	&Xupdate_avx_16_31(\&body_00_19);
1397	&Xupdate_avx_16_31(\&body_00_19);
1398	&Xupdate_avx_16_31(\&body_00_19);
1399	&Xupdate_avx_16_31(\&body_00_19);
1400	&Xupdate_avx_32_79(\&body_00_19);
1401	&Xupdate_avx_32_79(\&body_20_39);
1402	&Xupdate_avx_32_79(\&body_20_39);
1403	&Xupdate_avx_32_79(\&body_20_39);
1404	&Xupdate_avx_32_79(\&body_20_39);
1405	&Xupdate_avx_32_79(\&body_20_39);
1406	&Xupdate_avx_32_79(\&body_40_59);
1407	&Xupdate_avx_32_79(\&body_40_59);
1408	&Xupdate_avx_32_79(\&body_40_59);
1409	&Xupdate_avx_32_79(\&body_40_59);
1410	&Xupdate_avx_32_79(\&body_40_59);
1411	&Xupdate_avx_32_79(\&body_20_39);
1412	&Xuplast_avx_80(\&body_20_39);	# can jump to "done"
1413
1414				$saved_j=$j; @saved_V=@V;
1415
1416	&Xloop_avx(\&body_20_39);
1417	&Xloop_avx(\&body_20_39);
1418	&Xloop_avx(\&body_20_39);
1419
1420	&mov	(@T[1],&DWP(192,"esp"));	# update context
1421	&add	($A,&DWP(0,@T[1]));
1422	&add	(@T[0],&DWP(4,@T[1]));		# $b
1423	&add	($C,&DWP(8,@T[1]));
1424	&mov	(&DWP(0,@T[1]),$A);
1425	&add	($D,&DWP(12,@T[1]));
1426	&mov	(&DWP(4,@T[1]),@T[0]);
1427	&add	($E,&DWP(16,@T[1]));
1428	&mov	($B,$C);
1429	&mov	(&DWP(8,@T[1]),$C);
1430	&xor	($B,$D);
1431	&mov	(&DWP(12,@T[1]),$D);
1432	&mov	(&DWP(16,@T[1]),$E);
1433	&mov	(@T[1],@T[0]);
1434	&and	(@T[0],$B);
1435	&mov	($B,@T[1]);
1436
1437	&jmp	(&label("loop"));
1438
1439&set_label("done",16);		$j=$saved_j; @V=@saved_V;
1440
1441	&Xtail_avx(\&body_20_39);
1442	&Xtail_avx(\&body_20_39);
1443	&Xtail_avx(\&body_20_39);
1444
1445	&vzeroall();
1446
1447	&mov	(@T[1],&DWP(192,"esp"));	# update context
1448	&add	($A,&DWP(0,@T[1]));
1449	&mov	("esp",&DWP(192+12,"esp"));	# restore %esp
1450	&add	(@T[0],&DWP(4,@T[1]));		# $b
1451	&add	($C,&DWP(8,@T[1]));
1452	&mov	(&DWP(0,@T[1]),$A);
1453	&add	($D,&DWP(12,@T[1]));
1454	&mov	(&DWP(4,@T[1]),@T[0]);
1455	&add	($E,&DWP(16,@T[1]));
1456	&mov	(&DWP(8,@T[1]),$C);
1457	&mov	(&DWP(12,@T[1]),$D);
1458	&mov	(&DWP(16,@T[1]),$E);
1459&function_end("_sha1_block_data_order_avx");
1460}
1461&set_label("K_XX_XX",64);
1462&data_word(0x5a827999,0x5a827999,0x5a827999,0x5a827999);	# K_00_19
1463&data_word(0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1);	# K_20_39
1464&data_word(0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc);	# K_40_59
1465&data_word(0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6);	# K_60_79
1466&data_word(0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f);	# pbswap mask
1467&data_byte(0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0);
1468}
1469&asciz("SHA1 block transform for x86, CRYPTOGAMS by <appro\@openssl.org>");
1470
1471&asm_finish();
1472