• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1#!/usr/bin/env perl
2#
3# ====================================================================
4# 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# March, May, June 2010
11#
12# The module implements "4-bit" GCM GHASH function and underlying
13# single multiplication operation in GF(2^128). "4-bit" means that it
14# uses 256 bytes per-key table [+64/128 bytes fixed table]. It has two
15# code paths: vanilla x86 and vanilla MMX. Former will be executed on
16# 486 and Pentium, latter on all others. MMX GHASH features so called
17# "528B" variant of "4-bit" method utilizing additional 256+16 bytes
18# of per-key storage [+512 bytes shared table]. Performance results
19# are for streamed GHASH subroutine and are expressed in cycles per
20# processed byte, less is better:
21#
22#		gcc 2.95.3(*)	MMX assembler	x86 assembler
23#
24# Pentium	105/111(**)	-		50
25# PIII		68 /75		12.2		24
26# P4		125/125		17.8		84(***)
27# Opteron	66 /70		10.1		30
28# Core2		54 /67		8.4		18
29#
30# (*)	gcc 3.4.x was observed to generate few percent slower code,
31#	which is one of reasons why 2.95.3 results were chosen,
32#	another reason is lack of 3.4.x results for older CPUs;
33#	comparison with MMX results is not completely fair, because C
34#	results are for vanilla "256B" implementation, while
35#	assembler results are for "528B";-)
36# (**)	second number is result for code compiled with -fPIC flag,
37#	which is actually more relevant, because assembler code is
38#	position-independent;
39# (***)	see comment in non-MMX routine for further details;
40#
41# To summarize, it's >2-5 times faster than gcc-generated code. To
42# anchor it to something else SHA1 assembler processes one byte in
43# 11-13 cycles on contemporary x86 cores. As for choice of MMX in
44# particular, see comment at the end of the file...
45
46# May 2010
47#
48# Add PCLMULQDQ version performing at 2.10 cycles per processed byte.
49# The question is how close is it to theoretical limit? The pclmulqdq
50# instruction latency appears to be 14 cycles and there can't be more
51# than 2 of them executing at any given time. This means that single
52# Karatsuba multiplication would take 28 cycles *plus* few cycles for
53# pre- and post-processing. Then multiplication has to be followed by
54# modulo-reduction. Given that aggregated reduction method [see
55# "Carry-less Multiplication and Its Usage for Computing the GCM Mode"
56# white paper by Intel] allows you to perform reduction only once in
57# a while we can assume that asymptotic performance can be estimated
58# as (28+Tmod/Naggr)/16, where Tmod is time to perform reduction
59# and Naggr is the aggregation factor.
60#
61# Before we proceed to this implementation let's have closer look at
62# the best-performing code suggested by Intel in their white paper.
63# By tracing inter-register dependencies Tmod is estimated as ~19
64# cycles and Naggr chosen by Intel is 4, resulting in 2.05 cycles per
65# processed byte. As implied, this is quite optimistic estimate,
66# because it does not account for Karatsuba pre- and post-processing,
67# which for a single multiplication is ~5 cycles. Unfortunately Intel
68# does not provide performance data for GHASH alone. But benchmarking
69# AES_GCM_encrypt ripped out of Fig. 15 of the white paper with aadt
70# alone resulted in 2.46 cycles per byte of out 16KB buffer. Note that
71# the result accounts even for pre-computing of degrees of the hash
72# key H, but its portion is negligible at 16KB buffer size.
73#
74# Moving on to the implementation in question. Tmod is estimated as
75# ~13 cycles and Naggr is 2, giving asymptotic performance of ...
76# 2.16. How is it possible that measured performance is better than
77# optimistic theoretical estimate? There is one thing Intel failed
78# to recognize. By serializing GHASH with CTR in same subroutine
79# former's performance is really limited to above (Tmul + Tmod/Naggr)
80# equation. But if GHASH procedure is detached, the modulo-reduction
81# can be interleaved with Naggr-1 multiplications at instruction level
82# and under ideal conditions even disappear from the equation. So that
83# optimistic theoretical estimate for this implementation is ...
84# 28/16=1.75, and not 2.16. Well, it's probably way too optimistic,
85# at least for such small Naggr. I'd argue that (28+Tproc/Naggr),
86# where Tproc is time required for Karatsuba pre- and post-processing,
87# is more realistic estimate. In this case it gives ... 1.91 cycles.
88# Or in other words, depending on how well we can interleave reduction
89# and one of the two multiplications the performance should be betwen
90# 1.91 and 2.16. As already mentioned, this implementation processes
91# one byte out of 8KB buffer in 2.10 cycles, while x86_64 counterpart
92# - in 2.02. x86_64 performance is better, because larger register
93# bank allows to interleave reduction and multiplication better.
94#
95# Does it make sense to increase Naggr? To start with it's virtually
96# impossible in 32-bit mode, because of limited register bank
97# capacity. Otherwise improvement has to be weighed agiainst slower
98# setup, as well as code size and complexity increase. As even
99# optimistic estimate doesn't promise 30% performance improvement,
100# there are currently no plans to increase Naggr.
101#
102# Special thanks to David Woodhouse <dwmw2@infradead.org> for
103# providing access to a Westmere-based system on behalf of Intel
104# Open Source Technology Centre.
105
106# January 2010
107#
108# Tweaked to optimize transitions between integer and FP operations
109# on same XMM register, PCLMULQDQ subroutine was measured to process
110# one byte in 2.07 cycles on Sandy Bridge, and in 2.12 - on Westmere.
111# The minor regression on Westmere is outweighed by ~15% improvement
112# on Sandy Bridge. Strangely enough attempt to modify 64-bit code in
113# similar manner resulted in almost 20% degradation on Sandy Bridge,
114# where original 64-bit code processes one byte in 1.95 cycles.
115
116$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
117push(@INC,"${dir}","${dir}../../perlasm");
118require "x86asm.pl";
119
120&asm_init($ARGV[0],"ghash-x86.pl",$x86only = $ARGV[$#ARGV] eq "386");
121
122$sse2=0;
123for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
124
125($Zhh,$Zhl,$Zlh,$Zll) = ("ebp","edx","ecx","ebx");
126$inp  = "edi";
127$Htbl = "esi";
128
129$unroll = 0;	# Affects x86 loop. Folded loop performs ~7% worse
130		# than unrolled, which has to be weighted against
131		# 2.5x x86-specific code size reduction.
132
133sub x86_loop {
134    my $off = shift;
135    my $rem = "eax";
136
137	&mov	($Zhh,&DWP(4,$Htbl,$Zll));
138	&mov	($Zhl,&DWP(0,$Htbl,$Zll));
139	&mov	($Zlh,&DWP(12,$Htbl,$Zll));
140	&mov	($Zll,&DWP(8,$Htbl,$Zll));
141	&xor	($rem,$rem);	# avoid partial register stalls on PIII
142
143	# shrd practically kills P4, 2.5x deterioration, but P4 has
144	# MMX code-path to execute. shrd runs tad faster [than twice
145	# the shifts, move's and or's] on pre-MMX Pentium (as well as
146	# PIII and Core2), *but* minimizes code size, spares register
147	# and thus allows to fold the loop...
148	if (!$unroll) {
149	my $cnt = $inp;
150	&mov	($cnt,15);
151	&jmp	(&label("x86_loop"));
152	&set_label("x86_loop",16);
153	    for($i=1;$i<=2;$i++) {
154		&mov	(&LB($rem),&LB($Zll));
155		&shrd	($Zll,$Zlh,4);
156		&and	(&LB($rem),0xf);
157		&shrd	($Zlh,$Zhl,4);
158		&shrd	($Zhl,$Zhh,4);
159		&shr	($Zhh,4);
160		&xor	($Zhh,&DWP($off+16,"esp",$rem,4));
161
162		&mov	(&LB($rem),&BP($off,"esp",$cnt));
163		if ($i&1) {
164			&and	(&LB($rem),0xf0);
165		} else {
166			&shl	(&LB($rem),4);
167		}
168
169		&xor	($Zll,&DWP(8,$Htbl,$rem));
170		&xor	($Zlh,&DWP(12,$Htbl,$rem));
171		&xor	($Zhl,&DWP(0,$Htbl,$rem));
172		&xor	($Zhh,&DWP(4,$Htbl,$rem));
173
174		if ($i&1) {
175			&dec	($cnt);
176			&js	(&label("x86_break"));
177		} else {
178			&jmp	(&label("x86_loop"));
179		}
180	    }
181	&set_label("x86_break",16);
182	} else {
183	    for($i=1;$i<32;$i++) {
184		&comment($i);
185		&mov	(&LB($rem),&LB($Zll));
186		&shrd	($Zll,$Zlh,4);
187		&and	(&LB($rem),0xf);
188		&shrd	($Zlh,$Zhl,4);
189		&shrd	($Zhl,$Zhh,4);
190		&shr	($Zhh,4);
191		&xor	($Zhh,&DWP($off+16,"esp",$rem,4));
192
193		if ($i&1) {
194			&mov	(&LB($rem),&BP($off+15-($i>>1),"esp"));
195			&and	(&LB($rem),0xf0);
196		} else {
197			&mov	(&LB($rem),&BP($off+15-($i>>1),"esp"));
198			&shl	(&LB($rem),4);
199		}
200
201		&xor	($Zll,&DWP(8,$Htbl,$rem));
202		&xor	($Zlh,&DWP(12,$Htbl,$rem));
203		&xor	($Zhl,&DWP(0,$Htbl,$rem));
204		&xor	($Zhh,&DWP(4,$Htbl,$rem));
205	    }
206	}
207	&bswap	($Zll);
208	&bswap	($Zlh);
209	&bswap	($Zhl);
210	if (!$x86only) {
211		&bswap	($Zhh);
212	} else {
213		&mov	("eax",$Zhh);
214		&bswap	("eax");
215		&mov	($Zhh,"eax");
216	}
217}
218
219if ($unroll) {
220    &function_begin_B("_x86_gmult_4bit_inner");
221	&x86_loop(4);
222	&ret	();
223    &function_end_B("_x86_gmult_4bit_inner");
224}
225
226sub deposit_rem_4bit {
227    my $bias = shift;
228
229	&mov	(&DWP($bias+0, "esp"),0x0000<<16);
230	&mov	(&DWP($bias+4, "esp"),0x1C20<<16);
231	&mov	(&DWP($bias+8, "esp"),0x3840<<16);
232	&mov	(&DWP($bias+12,"esp"),0x2460<<16);
233	&mov	(&DWP($bias+16,"esp"),0x7080<<16);
234	&mov	(&DWP($bias+20,"esp"),0x6CA0<<16);
235	&mov	(&DWP($bias+24,"esp"),0x48C0<<16);
236	&mov	(&DWP($bias+28,"esp"),0x54E0<<16);
237	&mov	(&DWP($bias+32,"esp"),0xE100<<16);
238	&mov	(&DWP($bias+36,"esp"),0xFD20<<16);
239	&mov	(&DWP($bias+40,"esp"),0xD940<<16);
240	&mov	(&DWP($bias+44,"esp"),0xC560<<16);
241	&mov	(&DWP($bias+48,"esp"),0x9180<<16);
242	&mov	(&DWP($bias+52,"esp"),0x8DA0<<16);
243	&mov	(&DWP($bias+56,"esp"),0xA9C0<<16);
244	&mov	(&DWP($bias+60,"esp"),0xB5E0<<16);
245}
246
247$suffix = $x86only ? "" : "_x86";
248
249&function_begin("gcm_gmult_4bit".$suffix);
250	&stack_push(16+4+1);			# +1 for stack alignment
251	&mov	($inp,&wparam(0));		# load Xi
252	&mov	($Htbl,&wparam(1));		# load Htable
253
254	&mov	($Zhh,&DWP(0,$inp));		# load Xi[16]
255	&mov	($Zhl,&DWP(4,$inp));
256	&mov	($Zlh,&DWP(8,$inp));
257	&mov	($Zll,&DWP(12,$inp));
258
259	&deposit_rem_4bit(16);
260
261	&mov	(&DWP(0,"esp"),$Zhh);		# copy Xi[16] on stack
262	&mov	(&DWP(4,"esp"),$Zhl);
263	&mov	(&DWP(8,"esp"),$Zlh);
264	&mov	(&DWP(12,"esp"),$Zll);
265	&shr	($Zll,20);
266	&and	($Zll,0xf0);
267
268	if ($unroll) {
269		&call	("_x86_gmult_4bit_inner");
270	} else {
271		&x86_loop(0);
272		&mov	($inp,&wparam(0));
273	}
274
275	&mov	(&DWP(12,$inp),$Zll);
276	&mov	(&DWP(8,$inp),$Zlh);
277	&mov	(&DWP(4,$inp),$Zhl);
278	&mov	(&DWP(0,$inp),$Zhh);
279	&stack_pop(16+4+1);
280&function_end("gcm_gmult_4bit".$suffix);
281
282&function_begin("gcm_ghash_4bit".$suffix);
283	&stack_push(16+4+1);			# +1 for 64-bit alignment
284	&mov	($Zll,&wparam(0));		# load Xi
285	&mov	($Htbl,&wparam(1));		# load Htable
286	&mov	($inp,&wparam(2));		# load in
287	&mov	("ecx",&wparam(3));		# load len
288	&add	("ecx",$inp);
289	&mov	(&wparam(3),"ecx");
290
291	&mov	($Zhh,&DWP(0,$Zll));		# load Xi[16]
292	&mov	($Zhl,&DWP(4,$Zll));
293	&mov	($Zlh,&DWP(8,$Zll));
294	&mov	($Zll,&DWP(12,$Zll));
295
296	&deposit_rem_4bit(16);
297
298    &set_label("x86_outer_loop",16);
299	&xor	($Zll,&DWP(12,$inp));		# xor with input
300	&xor	($Zlh,&DWP(8,$inp));
301	&xor	($Zhl,&DWP(4,$inp));
302	&xor	($Zhh,&DWP(0,$inp));
303	&mov	(&DWP(12,"esp"),$Zll);		# dump it on stack
304	&mov	(&DWP(8,"esp"),$Zlh);
305	&mov	(&DWP(4,"esp"),$Zhl);
306	&mov	(&DWP(0,"esp"),$Zhh);
307
308	&shr	($Zll,20);
309	&and	($Zll,0xf0);
310
311	if ($unroll) {
312		&call	("_x86_gmult_4bit_inner");
313	} else {
314		&x86_loop(0);
315		&mov	($inp,&wparam(2));
316	}
317	&lea	($inp,&DWP(16,$inp));
318	&cmp	($inp,&wparam(3));
319	&mov	(&wparam(2),$inp)	if (!$unroll);
320	&jb	(&label("x86_outer_loop"));
321
322	&mov	($inp,&wparam(0));	# load Xi
323	&mov	(&DWP(12,$inp),$Zll);
324	&mov	(&DWP(8,$inp),$Zlh);
325	&mov	(&DWP(4,$inp),$Zhl);
326	&mov	(&DWP(0,$inp),$Zhh);
327	&stack_pop(16+4+1);
328&function_end("gcm_ghash_4bit".$suffix);
329
330if (!$x86only) {{{
331
332&static_label("rem_4bit");
333
334if (!$sse2) {{	# pure-MMX "May" version...
335
336$S=12;		# shift factor for rem_4bit
337
338&function_begin_B("_mmx_gmult_4bit_inner");
339# MMX version performs 3.5 times better on P4 (see comment in non-MMX
340# routine for further details), 100% better on Opteron, ~70% better
341# on Core2 and PIII... In other words effort is considered to be well
342# spent... Since initial release the loop was unrolled in order to
343# "liberate" register previously used as loop counter. Instead it's
344# used to optimize critical path in 'Z.hi ^= rem_4bit[Z.lo&0xf]'.
345# The path involves move of Z.lo from MMX to integer register,
346# effective address calculation and finally merge of value to Z.hi.
347# Reference to rem_4bit is scheduled so late that I had to >>4
348# rem_4bit elements. This resulted in 20-45% procent improvement
349# on contemporary µ-archs.
350{
351    my $cnt;
352    my $rem_4bit = "eax";
353    my @rem = ($Zhh,$Zll);
354    my $nhi = $Zhl;
355    my $nlo = $Zlh;
356
357    my ($Zlo,$Zhi) = ("mm0","mm1");
358    my $tmp = "mm2";
359
360	&xor	($nlo,$nlo);	# avoid partial register stalls on PIII
361	&mov	($nhi,$Zll);
362	&mov	(&LB($nlo),&LB($nhi));
363	&shl	(&LB($nlo),4);
364	&and	($nhi,0xf0);
365	&movq	($Zlo,&QWP(8,$Htbl,$nlo));
366	&movq	($Zhi,&QWP(0,$Htbl,$nlo));
367	&movd	($rem[0],$Zlo);
368
369	for ($cnt=28;$cnt>=-2;$cnt--) {
370	    my $odd = $cnt&1;
371	    my $nix = $odd ? $nlo : $nhi;
372
373		&shl	(&LB($nlo),4)			if ($odd);
374		&psrlq	($Zlo,4);
375		&movq	($tmp,$Zhi);
376		&psrlq	($Zhi,4);
377		&pxor	($Zlo,&QWP(8,$Htbl,$nix));
378		&mov	(&LB($nlo),&BP($cnt/2,$inp))	if (!$odd && $cnt>=0);
379		&psllq	($tmp,60);
380		&and	($nhi,0xf0)			if ($odd);
381		&pxor	($Zhi,&QWP(0,$rem_4bit,$rem[1],8)) if ($cnt<28);
382		&and	($rem[0],0xf);
383		&pxor	($Zhi,&QWP(0,$Htbl,$nix));
384		&mov	($nhi,$nlo)			if (!$odd && $cnt>=0);
385		&movd	($rem[1],$Zlo);
386		&pxor	($Zlo,$tmp);
387
388		push	(@rem,shift(@rem));		# "rotate" registers
389	}
390
391	&mov	($inp,&DWP(4,$rem_4bit,$rem[1],8));	# last rem_4bit[rem]
392
393	&psrlq	($Zlo,32);	# lower part of Zlo is already there
394	&movd	($Zhl,$Zhi);
395	&psrlq	($Zhi,32);
396	&movd	($Zlh,$Zlo);
397	&movd	($Zhh,$Zhi);
398	&shl	($inp,4);	# compensate for rem_4bit[i] being >>4
399
400	&bswap	($Zll);
401	&bswap	($Zhl);
402	&bswap	($Zlh);
403	&xor	($Zhh,$inp);
404	&bswap	($Zhh);
405
406	&ret	();
407}
408&function_end_B("_mmx_gmult_4bit_inner");
409
410&function_begin("gcm_gmult_4bit_mmx");
411	&mov	($inp,&wparam(0));	# load Xi
412	&mov	($Htbl,&wparam(1));	# load Htable
413
414	&call	(&label("pic_point"));
415	&set_label("pic_point");
416	&blindpop("eax");
417	&lea	("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
418
419	&movz	($Zll,&BP(15,$inp));
420
421	&call	("_mmx_gmult_4bit_inner");
422
423	&mov	($inp,&wparam(0));	# load Xi
424	&emms	();
425	&mov	(&DWP(12,$inp),$Zll);
426	&mov	(&DWP(4,$inp),$Zhl);
427	&mov	(&DWP(8,$inp),$Zlh);
428	&mov	(&DWP(0,$inp),$Zhh);
429&function_end("gcm_gmult_4bit_mmx");
430
431# Streamed version performs 20% better on P4, 7% on Opteron,
432# 10% on Core2 and PIII...
433&function_begin("gcm_ghash_4bit_mmx");
434	&mov	($Zhh,&wparam(0));	# load Xi
435	&mov	($Htbl,&wparam(1));	# load Htable
436	&mov	($inp,&wparam(2));	# load in
437	&mov	($Zlh,&wparam(3));	# load len
438
439	&call	(&label("pic_point"));
440	&set_label("pic_point");
441	&blindpop("eax");
442	&lea	("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
443
444	&add	($Zlh,$inp);
445	&mov	(&wparam(3),$Zlh);	# len to point at the end of input
446	&stack_push(4+1);		# +1 for stack alignment
447
448	&mov	($Zll,&DWP(12,$Zhh));	# load Xi[16]
449	&mov	($Zhl,&DWP(4,$Zhh));
450	&mov	($Zlh,&DWP(8,$Zhh));
451	&mov	($Zhh,&DWP(0,$Zhh));
452	&jmp	(&label("mmx_outer_loop"));
453
454    &set_label("mmx_outer_loop",16);
455	&xor	($Zll,&DWP(12,$inp));
456	&xor	($Zhl,&DWP(4,$inp));
457	&xor	($Zlh,&DWP(8,$inp));
458	&xor	($Zhh,&DWP(0,$inp));
459	&mov	(&wparam(2),$inp);
460	&mov	(&DWP(12,"esp"),$Zll);
461	&mov	(&DWP(4,"esp"),$Zhl);
462	&mov	(&DWP(8,"esp"),$Zlh);
463	&mov	(&DWP(0,"esp"),$Zhh);
464
465	&mov	($inp,"esp");
466	&shr	($Zll,24);
467
468	&call	("_mmx_gmult_4bit_inner");
469
470	&mov	($inp,&wparam(2));
471	&lea	($inp,&DWP(16,$inp));
472	&cmp	($inp,&wparam(3));
473	&jb	(&label("mmx_outer_loop"));
474
475	&mov	($inp,&wparam(0));	# load Xi
476	&emms	();
477	&mov	(&DWP(12,$inp),$Zll);
478	&mov	(&DWP(4,$inp),$Zhl);
479	&mov	(&DWP(8,$inp),$Zlh);
480	&mov	(&DWP(0,$inp),$Zhh);
481
482	&stack_pop(4+1);
483&function_end("gcm_ghash_4bit_mmx");
484
485}} else {{	# "June" MMX version...
486		# ... has slower "April" gcm_gmult_4bit_mmx with folded
487		# loop. This is done to conserve code size...
488$S=16;		# shift factor for rem_4bit
489
490sub mmx_loop() {
491# MMX version performs 2.8 times better on P4 (see comment in non-MMX
492# routine for further details), 40% better on Opteron and Core2, 50%
493# better on PIII... In other words effort is considered to be well
494# spent...
495    my $inp = shift;
496    my $rem_4bit = shift;
497    my $cnt = $Zhh;
498    my $nhi = $Zhl;
499    my $nlo = $Zlh;
500    my $rem = $Zll;
501
502    my ($Zlo,$Zhi) = ("mm0","mm1");
503    my $tmp = "mm2";
504
505	&xor	($nlo,$nlo);	# avoid partial register stalls on PIII
506	&mov	($nhi,$Zll);
507	&mov	(&LB($nlo),&LB($nhi));
508	&mov	($cnt,14);
509	&shl	(&LB($nlo),4);
510	&and	($nhi,0xf0);
511	&movq	($Zlo,&QWP(8,$Htbl,$nlo));
512	&movq	($Zhi,&QWP(0,$Htbl,$nlo));
513	&movd	($rem,$Zlo);
514	&jmp	(&label("mmx_loop"));
515
516    &set_label("mmx_loop",16);
517	&psrlq	($Zlo,4);
518	&and	($rem,0xf);
519	&movq	($tmp,$Zhi);
520	&psrlq	($Zhi,4);
521	&pxor	($Zlo,&QWP(8,$Htbl,$nhi));
522	&mov	(&LB($nlo),&BP(0,$inp,$cnt));
523	&psllq	($tmp,60);
524	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
525	&dec	($cnt);
526	&movd	($rem,$Zlo);
527	&pxor	($Zhi,&QWP(0,$Htbl,$nhi));
528	&mov	($nhi,$nlo);
529	&pxor	($Zlo,$tmp);
530	&js	(&label("mmx_break"));
531
532	&shl	(&LB($nlo),4);
533	&and	($rem,0xf);
534	&psrlq	($Zlo,4);
535	&and	($nhi,0xf0);
536	&movq	($tmp,$Zhi);
537	&psrlq	($Zhi,4);
538	&pxor	($Zlo,&QWP(8,$Htbl,$nlo));
539	&psllq	($tmp,60);
540	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
541	&movd	($rem,$Zlo);
542	&pxor	($Zhi,&QWP(0,$Htbl,$nlo));
543	&pxor	($Zlo,$tmp);
544	&jmp	(&label("mmx_loop"));
545
546    &set_label("mmx_break",16);
547	&shl	(&LB($nlo),4);
548	&and	($rem,0xf);
549	&psrlq	($Zlo,4);
550	&and	($nhi,0xf0);
551	&movq	($tmp,$Zhi);
552	&psrlq	($Zhi,4);
553	&pxor	($Zlo,&QWP(8,$Htbl,$nlo));
554	&psllq	($tmp,60);
555	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
556	&movd	($rem,$Zlo);
557	&pxor	($Zhi,&QWP(0,$Htbl,$nlo));
558	&pxor	($Zlo,$tmp);
559
560	&psrlq	($Zlo,4);
561	&and	($rem,0xf);
562	&movq	($tmp,$Zhi);
563	&psrlq	($Zhi,4);
564	&pxor	($Zlo,&QWP(8,$Htbl,$nhi));
565	&psllq	($tmp,60);
566	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
567	&movd	($rem,$Zlo);
568	&pxor	($Zhi,&QWP(0,$Htbl,$nhi));
569	&pxor	($Zlo,$tmp);
570
571	&psrlq	($Zlo,32);	# lower part of Zlo is already there
572	&movd	($Zhl,$Zhi);
573	&psrlq	($Zhi,32);
574	&movd	($Zlh,$Zlo);
575	&movd	($Zhh,$Zhi);
576
577	&bswap	($Zll);
578	&bswap	($Zhl);
579	&bswap	($Zlh);
580	&bswap	($Zhh);
581}
582
583&function_begin("gcm_gmult_4bit_mmx");
584	&mov	($inp,&wparam(0));	# load Xi
585	&mov	($Htbl,&wparam(1));	# load Htable
586
587	&call	(&label("pic_point"));
588	&set_label("pic_point");
589	&blindpop("eax");
590	&lea	("eax",&DWP(&label("rem_4bit")."-".&label("pic_point"),"eax"));
591
592	&movz	($Zll,&BP(15,$inp));
593
594	&mmx_loop($inp,"eax");
595
596	&emms	();
597	&mov	(&DWP(12,$inp),$Zll);
598	&mov	(&DWP(4,$inp),$Zhl);
599	&mov	(&DWP(8,$inp),$Zlh);
600	&mov	(&DWP(0,$inp),$Zhh);
601&function_end("gcm_gmult_4bit_mmx");
602
603######################################################################
604# Below subroutine is "528B" variant of "4-bit" GCM GHASH function
605# (see gcm128.c for details). It provides further 20-40% performance
606# improvement over above mentioned "May" version.
607
608&static_label("rem_8bit");
609
610&function_begin("gcm_ghash_4bit_mmx");
611{ my ($Zlo,$Zhi) = ("mm7","mm6");
612  my $rem_8bit = "esi";
613  my $Htbl = "ebx";
614
615    # parameter block
616    &mov	("eax",&wparam(0));		# Xi
617    &mov	("ebx",&wparam(1));		# Htable
618    &mov	("ecx",&wparam(2));		# inp
619    &mov	("edx",&wparam(3));		# len
620    &mov	("ebp","esp");			# original %esp
621    &call	(&label("pic_point"));
622    &set_label	("pic_point");
623    &blindpop	($rem_8bit);
624    &lea	($rem_8bit,&DWP(&label("rem_8bit")."-".&label("pic_point"),$rem_8bit));
625
626    &sub	("esp",512+16+16);		# allocate stack frame...
627    &and	("esp",-64);			# ...and align it
628    &sub	("esp",16);			# place for (u8)(H[]<<4)
629
630    &add	("edx","ecx");			# pointer to the end of input
631    &mov	(&DWP(528+16+0,"esp"),"eax");	# save Xi
632    &mov	(&DWP(528+16+8,"esp"),"edx");	# save inp+len
633    &mov	(&DWP(528+16+12,"esp"),"ebp");	# save original %esp
634
635    { my @lo  = ("mm0","mm1","mm2");
636      my @hi  = ("mm3","mm4","mm5");
637      my @tmp = ("mm6","mm7");
638      my ($off1,$off2,$i) = (0,0,);
639
640      &add	($Htbl,128);			# optimize for size
641      &lea	("edi",&DWP(16+128,"esp"));
642      &lea	("ebp",&DWP(16+256+128,"esp"));
643
644      # decompose Htable (low and high parts are kept separately),
645      # generate Htable[]>>4, (u8)(Htable[]<<4), save to stack...
646      for ($i=0;$i<18;$i++) {
647
648	&mov	("edx",&DWP(16*$i+8-128,$Htbl))		if ($i<16);
649	&movq	($lo[0],&QWP(16*$i+8-128,$Htbl))	if ($i<16);
650	&psllq	($tmp[1],60)				if ($i>1);
651	&movq	($hi[0],&QWP(16*$i+0-128,$Htbl))	if ($i<16);
652	&por	($lo[2],$tmp[1])			if ($i>1);
653	&movq	(&QWP($off1-128,"edi"),$lo[1])		if ($i>0 && $i<17);
654	&psrlq	($lo[1],4)				if ($i>0 && $i<17);
655	&movq	(&QWP($off1,"edi"),$hi[1])		if ($i>0 && $i<17);
656	&movq	($tmp[0],$hi[1])			if ($i>0 && $i<17);
657	&movq	(&QWP($off2-128,"ebp"),$lo[2])		if ($i>1);
658	&psrlq	($hi[1],4)				if ($i>0 && $i<17);
659	&movq	(&QWP($off2,"ebp"),$hi[2])		if ($i>1);
660	&shl	("edx",4)				if ($i<16);
661	&mov	(&BP($i,"esp"),&LB("edx"))		if ($i<16);
662
663	unshift	(@lo,pop(@lo));			# "rotate" registers
664	unshift	(@hi,pop(@hi));
665	unshift	(@tmp,pop(@tmp));
666	$off1 += 8	if ($i>0);
667	$off2 += 8	if ($i>1);
668      }
669    }
670
671    &movq	($Zhi,&QWP(0,"eax"));
672    &mov	("ebx",&DWP(8,"eax"));
673    &mov	("edx",&DWP(12,"eax"));		# load Xi
674
675&set_label("outer",16);
676  { my $nlo = "eax";
677    my $dat = "edx";
678    my @nhi = ("edi","ebp");
679    my @rem = ("ebx","ecx");
680    my @red = ("mm0","mm1","mm2");
681    my $tmp = "mm3";
682
683    &xor	($dat,&DWP(12,"ecx"));		# merge input data
684    &xor	("ebx",&DWP(8,"ecx"));
685    &pxor	($Zhi,&QWP(0,"ecx"));
686    &lea	("ecx",&DWP(16,"ecx"));		# inp+=16
687    #&mov	(&DWP(528+12,"esp"),$dat);	# save inp^Xi
688    &mov	(&DWP(528+8,"esp"),"ebx");
689    &movq	(&QWP(528+0,"esp"),$Zhi);
690    &mov	(&DWP(528+16+4,"esp"),"ecx");	# save inp
691
692    &xor	($nlo,$nlo);
693    &rol	($dat,8);
694    &mov	(&LB($nlo),&LB($dat));
695    &mov	($nhi[1],$nlo);
696    &and	(&LB($nlo),0x0f);
697    &shr	($nhi[1],4);
698    &pxor	($red[0],$red[0]);
699    &rol	($dat,8);			# next byte
700    &pxor	($red[1],$red[1]);
701    &pxor	($red[2],$red[2]);
702
703    # Just like in "May" verson modulo-schedule for critical path in
704    # 'Z.hi ^= rem_8bit[Z.lo&0xff^((u8)H[nhi]<<4)]<<48'. Final 'pxor'
705    # is scheduled so late that rem_8bit[] has to be shifted *right*
706    # by 16, which is why last argument to pinsrw is 2, which
707    # corresponds to <<32=<<48>>16...
708    for ($j=11,$i=0;$i<15;$i++) {
709
710      if ($i>0) {
711	&pxor	($Zlo,&QWP(16,"esp",$nlo,8));		# Z^=H[nlo]
712	&rol	($dat,8);				# next byte
713	&pxor	($Zhi,&QWP(16+128,"esp",$nlo,8));
714
715	&pxor	($Zlo,$tmp);
716	&pxor	($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
717	&xor	(&LB($rem[1]),&BP(0,"esp",$nhi[0]));	# rem^(H[nhi]<<4)
718      } else {
719	&movq	($Zlo,&QWP(16,"esp",$nlo,8));
720	&movq	($Zhi,&QWP(16+128,"esp",$nlo,8));
721      }
722
723	&mov	(&LB($nlo),&LB($dat));
724	&mov	($dat,&DWP(528+$j,"esp"))		if (--$j%4==0);
725
726	&movd	($rem[0],$Zlo);
727	&movz	($rem[1],&LB($rem[1]))			if ($i>0);
728	&psrlq	($Zlo,8);				# Z>>=8
729
730	&movq	($tmp,$Zhi);
731	&mov	($nhi[0],$nlo);
732	&psrlq	($Zhi,8);
733
734	&pxor	($Zlo,&QWP(16+256+0,"esp",$nhi[1],8));	# Z^=H[nhi]>>4
735	&and	(&LB($nlo),0x0f);
736	&psllq	($tmp,56);
737
738	&pxor	($Zhi,$red[1])				if ($i>1);
739	&shr	($nhi[0],4);
740	&pinsrw	($red[0],&WP(0,$rem_8bit,$rem[1],2),2)	if ($i>0);
741
742	unshift	(@red,pop(@red));			# "rotate" registers
743	unshift	(@rem,pop(@rem));
744	unshift	(@nhi,pop(@nhi));
745    }
746
747    &pxor	($Zlo,&QWP(16,"esp",$nlo,8));		# Z^=H[nlo]
748    &pxor	($Zhi,&QWP(16+128,"esp",$nlo,8));
749    &xor	(&LB($rem[1]),&BP(0,"esp",$nhi[0]));	# rem^(H[nhi]<<4)
750
751    &pxor	($Zlo,$tmp);
752    &pxor	($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
753    &movz	($rem[1],&LB($rem[1]));
754
755    &pxor	($red[2],$red[2]);			# clear 2nd word
756    &psllq	($red[1],4);
757
758    &movd	($rem[0],$Zlo);
759    &psrlq	($Zlo,4);				# Z>>=4
760
761    &movq	($tmp,$Zhi);
762    &psrlq	($Zhi,4);
763    &shl	($rem[0],4);				# rem<<4
764
765    &pxor	($Zlo,&QWP(16,"esp",$nhi[1],8));	# Z^=H[nhi]
766    &psllq	($tmp,60);
767    &movz	($rem[0],&LB($rem[0]));
768
769    &pxor	($Zlo,$tmp);
770    &pxor	($Zhi,&QWP(16+128,"esp",$nhi[1],8));
771
772    &pinsrw	($red[0],&WP(0,$rem_8bit,$rem[1],2),2);
773    &pxor	($Zhi,$red[1]);
774
775    &movd	($dat,$Zlo);
776    &pinsrw	($red[2],&WP(0,$rem_8bit,$rem[0],2),3);	# last is <<48
777
778    &psllq	($red[0],12);				# correct by <<16>>4
779    &pxor	($Zhi,$red[0]);
780    &psrlq	($Zlo,32);
781    &pxor	($Zhi,$red[2]);
782
783    &mov	("ecx",&DWP(528+16+4,"esp"));	# restore inp
784    &movd	("ebx",$Zlo);
785    &movq	($tmp,$Zhi);			# 01234567
786    &psllw	($Zhi,8);			# 1.3.5.7.
787    &psrlw	($tmp,8);			# .0.2.4.6
788    &por	($Zhi,$tmp);			# 10325476
789    &bswap	($dat);
790    &pshufw	($Zhi,$Zhi,0b00011011);		# 76543210
791    &bswap	("ebx");
792
793    &cmp	("ecx",&DWP(528+16+8,"esp"));	# are we done?
794    &jne	(&label("outer"));
795  }
796
797    &mov	("eax",&DWP(528+16+0,"esp"));	# restore Xi
798    &mov	(&DWP(12,"eax"),"edx");
799    &mov	(&DWP(8,"eax"),"ebx");
800    &movq	(&QWP(0,"eax"),$Zhi);
801
802    &mov	("esp",&DWP(528+16+12,"esp"));	# restore original %esp
803    &emms	();
804}
805&function_end("gcm_ghash_4bit_mmx");
806}}
807
808if ($sse2) {{
809######################################################################
810# PCLMULQDQ version.
811
812$Xip="eax";
813$Htbl="edx";
814$const="ecx";
815$inp="esi";
816$len="ebx";
817
818($Xi,$Xhi)=("xmm0","xmm1");	$Hkey="xmm2";
819($T1,$T2,$T3)=("xmm3","xmm4","xmm5");
820($Xn,$Xhn)=("xmm6","xmm7");
821
822&static_label("bswap");
823
824sub clmul64x64_T2 {	# minimal "register" pressure
825my ($Xhi,$Xi,$Hkey)=@_;
826
827	&movdqa		($Xhi,$Xi);		#
828	&pshufd		($T1,$Xi,0b01001110);
829	&pshufd		($T2,$Hkey,0b01001110);
830	&pxor		($T1,$Xi);		#
831	&pxor		($T2,$Hkey);
832
833	&pclmulqdq	($Xi,$Hkey,0x00);	#######
834	&pclmulqdq	($Xhi,$Hkey,0x11);	#######
835	&pclmulqdq	($T1,$T2,0x00);		#######
836	&xorps		($T1,$Xi);		#
837	&xorps		($T1,$Xhi);		#
838
839	&movdqa		($T2,$T1);		#
840	&psrldq		($T1,8);
841	&pslldq		($T2,8);		#
842	&pxor		($Xhi,$T1);
843	&pxor		($Xi,$T2);		#
844}
845
846sub clmul64x64_T3 {
847# Even though this subroutine offers visually better ILP, it
848# was empirically found to be a tad slower than above version.
849# At least in gcm_ghash_clmul context. But it's just as well,
850# because loop modulo-scheduling is possible only thanks to
851# minimized "register" pressure...
852my ($Xhi,$Xi,$Hkey)=@_;
853
854	&movdqa		($T1,$Xi);		#
855	&movdqa		($Xhi,$Xi);
856	&pclmulqdq	($Xi,$Hkey,0x00);	#######
857	&pclmulqdq	($Xhi,$Hkey,0x11);	#######
858	&pshufd		($T2,$T1,0b01001110);	#
859	&pshufd		($T3,$Hkey,0b01001110);
860	&pxor		($T2,$T1);		#
861	&pxor		($T3,$Hkey);
862	&pclmulqdq	($T2,$T3,0x00);		#######
863	&pxor		($T2,$Xi);		#
864	&pxor		($T2,$Xhi);		#
865
866	&movdqa		($T3,$T2);		#
867	&psrldq		($T2,8);
868	&pslldq		($T3,8);		#
869	&pxor		($Xhi,$T2);
870	&pxor		($Xi,$T3);		#
871}
872
873if (1) {		# Algorithm 9 with <<1 twist.
874			# Reduction is shorter and uses only two
875			# temporary registers, which makes it better
876			# candidate for interleaving with 64x64
877			# multiplication. Pre-modulo-scheduled loop
878			# was found to be ~20% faster than Algorithm 5
879			# below. Algorithm 9 was therefore chosen for
880			# further optimization...
881
882sub reduction_alg9 {	# 17/13 times faster than Intel version
883my ($Xhi,$Xi) = @_;
884
885	# 1st phase
886	&movdqa		($T1,$Xi);		#
887	&psllq		($Xi,1);
888	&pxor		($Xi,$T1);		#
889	&psllq		($Xi,5);		#
890	&pxor		($Xi,$T1);		#
891	&psllq		($Xi,57);		#
892	&movdqa		($T2,$Xi);		#
893	&pslldq		($Xi,8);
894	&psrldq		($T2,8);		#
895	&pxor		($Xi,$T1);
896	&pxor		($Xhi,$T2);		#
897
898	# 2nd phase
899	&movdqa		($T2,$Xi);
900	&psrlq		($Xi,5);
901	&pxor		($Xi,$T2);		#
902	&psrlq		($Xi,1);		#
903	&pxor		($Xi,$T2);		#
904	&pxor		($T2,$Xhi);
905	&psrlq		($Xi,1);		#
906	&pxor		($Xi,$T2);		#
907}
908
909&function_begin_B("gcm_init_clmul");
910	&mov		($Htbl,&wparam(0));
911	&mov		($Xip,&wparam(1));
912
913	&call		(&label("pic"));
914&set_label("pic");
915	&blindpop	($const);
916	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
917
918	&movdqu		($Hkey,&QWP(0,$Xip));
919	&pshufd		($Hkey,$Hkey,0b01001110);# dword swap
920
921	# <<1 twist
922	&pshufd		($T2,$Hkey,0b11111111);	# broadcast uppermost dword
923	&movdqa		($T1,$Hkey);
924	&psllq		($Hkey,1);
925	&pxor		($T3,$T3);		#
926	&psrlq		($T1,63);
927	&pcmpgtd	($T3,$T2);		# broadcast carry bit
928	&pslldq		($T1,8);
929	&por		($Hkey,$T1);		# H<<=1
930
931	# magic reduction
932	&pand		($T3,&QWP(16,$const));	# 0x1c2_polynomial
933	&pxor		($Hkey,$T3);		# if(carry) H^=0x1c2_polynomial
934
935	# calculate H^2
936	&movdqa		($Xi,$Hkey);
937	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
938	&reduction_alg9	($Xhi,$Xi);
939
940	&movdqu		(&QWP(0,$Htbl),$Hkey);	# save H
941	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
942
943	&ret		();
944&function_end_B("gcm_init_clmul");
945
946&function_begin_B("gcm_gmult_clmul");
947	&mov		($Xip,&wparam(0));
948	&mov		($Htbl,&wparam(1));
949
950	&call		(&label("pic"));
951&set_label("pic");
952	&blindpop	($const);
953	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
954
955	&movdqu		($Xi,&QWP(0,$Xip));
956	&movdqa		($T3,&QWP(0,$const));
957	&movups		($Hkey,&QWP(0,$Htbl));
958	&pshufb		($Xi,$T3);
959
960	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
961	&reduction_alg9	($Xhi,$Xi);
962
963	&pshufb		($Xi,$T3);
964	&movdqu		(&QWP(0,$Xip),$Xi);
965
966	&ret	();
967&function_end_B("gcm_gmult_clmul");
968
969&function_begin("gcm_ghash_clmul");
970	&mov		($Xip,&wparam(0));
971	&mov		($Htbl,&wparam(1));
972	&mov		($inp,&wparam(2));
973	&mov		($len,&wparam(3));
974
975	&call		(&label("pic"));
976&set_label("pic");
977	&blindpop	($const);
978	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
979
980	&movdqu		($Xi,&QWP(0,$Xip));
981	&movdqa		($T3,&QWP(0,$const));
982	&movdqu		($Hkey,&QWP(0,$Htbl));
983	&pshufb		($Xi,$T3);
984
985	&sub		($len,0x10);
986	&jz		(&label("odd_tail"));
987
988	#######
989	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
990	#	[(H*Ii+1) + (H*Xi+1)] mod P =
991	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
992	#
993	&movdqu		($T1,&QWP(0,$inp));	# Ii
994	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
995	&pshufb		($T1,$T3);
996	&pshufb		($Xn,$T3);
997	&pxor		($Xi,$T1);		# Ii+Xi
998
999	&clmul64x64_T2	($Xhn,$Xn,$Hkey);	# H*Ii+1
1000	&movups		($Hkey,&QWP(16,$Htbl));	# load H^2
1001
1002	&lea		($inp,&DWP(32,$inp));	# i+=2
1003	&sub		($len,0x20);
1004	&jbe		(&label("even_tail"));
1005
1006&set_label("mod_loop");
1007	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1008	&movdqu		($T1,&QWP(0,$inp));	# Ii
1009	&movups		($Hkey,&QWP(0,$Htbl));	# load H
1010
1011	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1012	&pxor		($Xhi,$Xhn);
1013
1014	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1015	&pshufb		($T1,$T3);
1016	&pshufb		($Xn,$T3);
1017
1018	&movdqa		($T3,$Xn);		#&clmul64x64_TX	($Xhn,$Xn,$Hkey); H*Ii+1
1019	&movdqa		($Xhn,$Xn);
1020	 &pxor		($Xhi,$T1);		# "Ii+Xi", consume early
1021
1022	  &movdqa	($T1,$Xi);		#&reduction_alg9($Xhi,$Xi); 1st phase
1023	  &psllq	($Xi,1);
1024	  &pxor		($Xi,$T1);		#
1025	  &psllq	($Xi,5);		#
1026	  &pxor		($Xi,$T1);		#
1027	&pclmulqdq	($Xn,$Hkey,0x00);	#######
1028	  &psllq	($Xi,57);		#
1029	  &movdqa	($T2,$Xi);		#
1030	  &pslldq	($Xi,8);
1031	  &psrldq	($T2,8);		#
1032	  &pxor		($Xi,$T1);
1033	&pshufd		($T1,$T3,0b01001110);
1034	  &pxor		($Xhi,$T2);		#
1035	&pxor		($T1,$T3);
1036	&pshufd		($T3,$Hkey,0b01001110);
1037	&pxor		($T3,$Hkey);		#
1038
1039	&pclmulqdq	($Xhn,$Hkey,0x11);	#######
1040	  &movdqa	($T2,$Xi);		# 2nd phase
1041	  &psrlq	($Xi,5);
1042	  &pxor		($Xi,$T2);		#
1043	  &psrlq	($Xi,1);		#
1044	  &pxor		($Xi,$T2);		#
1045	  &pxor		($T2,$Xhi);
1046	  &psrlq	($Xi,1);		#
1047	  &pxor		($Xi,$T2);		#
1048
1049	&pclmulqdq	($T1,$T3,0x00);		#######
1050	&movups		($Hkey,&QWP(16,$Htbl));	# load H^2
1051	&xorps		($T1,$Xn);		#
1052	&xorps		($T1,$Xhn);		#
1053
1054	&movdqa		($T3,$T1);		#
1055	&psrldq		($T1,8);
1056	&pslldq		($T3,8);		#
1057	&pxor		($Xhn,$T1);
1058	&pxor		($Xn,$T3);		#
1059	&movdqa		($T3,&QWP(0,$const));
1060
1061	&lea		($inp,&DWP(32,$inp));
1062	&sub		($len,0x20);
1063	&ja		(&label("mod_loop"));
1064
1065&set_label("even_tail");
1066	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1067
1068	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1069	&pxor		($Xhi,$Xhn);
1070
1071	&reduction_alg9	($Xhi,$Xi);
1072
1073	&test		($len,$len);
1074	&jnz		(&label("done"));
1075
1076	&movups		($Hkey,&QWP(0,$Htbl));	# load H
1077&set_label("odd_tail");
1078	&movdqu		($T1,&QWP(0,$inp));	# Ii
1079	&pshufb		($T1,$T3);
1080	&pxor		($Xi,$T1);		# Ii+Xi
1081
1082	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1083	&reduction_alg9	($Xhi,$Xi);
1084
1085&set_label("done");
1086	&pshufb		($Xi,$T3);
1087	&movdqu		(&QWP(0,$Xip),$Xi);
1088&function_end("gcm_ghash_clmul");
1089
1090} else {		# Algorith 5. Kept for reference purposes.
1091
1092sub reduction_alg5 {	# 19/16 times faster than Intel version
1093my ($Xhi,$Xi)=@_;
1094
1095	# <<1
1096	&movdqa		($T1,$Xi);		#
1097	&movdqa		($T2,$Xhi);
1098	&pslld		($Xi,1);
1099	&pslld		($Xhi,1);		#
1100	&psrld		($T1,31);
1101	&psrld		($T2,31);		#
1102	&movdqa		($T3,$T1);
1103	&pslldq		($T1,4);
1104	&psrldq		($T3,12);		#
1105	&pslldq		($T2,4);
1106	&por		($Xhi,$T3);		#
1107	&por		($Xi,$T1);
1108	&por		($Xhi,$T2);		#
1109
1110	# 1st phase
1111	&movdqa		($T1,$Xi);
1112	&movdqa		($T2,$Xi);
1113	&movdqa		($T3,$Xi);		#
1114	&pslld		($T1,31);
1115	&pslld		($T2,30);
1116	&pslld		($Xi,25);		#
1117	&pxor		($T1,$T2);
1118	&pxor		($T1,$Xi);		#
1119	&movdqa		($T2,$T1);		#
1120	&pslldq		($T1,12);
1121	&psrldq		($T2,4);		#
1122	&pxor		($T3,$T1);
1123
1124	# 2nd phase
1125	&pxor		($Xhi,$T3);		#
1126	&movdqa		($Xi,$T3);
1127	&movdqa		($T1,$T3);
1128	&psrld		($Xi,1);		#
1129	&psrld		($T1,2);
1130	&psrld		($T3,7);		#
1131	&pxor		($Xi,$T1);
1132	&pxor		($Xhi,$T2);
1133	&pxor		($Xi,$T3);		#
1134	&pxor		($Xi,$Xhi);		#
1135}
1136
1137&function_begin_B("gcm_init_clmul");
1138	&mov		($Htbl,&wparam(0));
1139	&mov		($Xip,&wparam(1));
1140
1141	&call		(&label("pic"));
1142&set_label("pic");
1143	&blindpop	($const);
1144	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1145
1146	&movdqu		($Hkey,&QWP(0,$Xip));
1147	&pshufd		($Hkey,$Hkey,0b01001110);# dword swap
1148
1149	# calculate H^2
1150	&movdqa		($Xi,$Hkey);
1151	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1152	&reduction_alg5	($Xhi,$Xi);
1153
1154	&movdqu		(&QWP(0,$Htbl),$Hkey);	# save H
1155	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
1156
1157	&ret		();
1158&function_end_B("gcm_init_clmul");
1159
1160&function_begin_B("gcm_gmult_clmul");
1161	&mov		($Xip,&wparam(0));
1162	&mov		($Htbl,&wparam(1));
1163
1164	&call		(&label("pic"));
1165&set_label("pic");
1166	&blindpop	($const);
1167	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1168
1169	&movdqu		($Xi,&QWP(0,$Xip));
1170	&movdqa		($Xn,&QWP(0,$const));
1171	&movdqu		($Hkey,&QWP(0,$Htbl));
1172	&pshufb		($Xi,$Xn);
1173
1174	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1175	&reduction_alg5	($Xhi,$Xi);
1176
1177	&pshufb		($Xi,$Xn);
1178	&movdqu		(&QWP(0,$Xip),$Xi);
1179
1180	&ret	();
1181&function_end_B("gcm_gmult_clmul");
1182
1183&function_begin("gcm_ghash_clmul");
1184	&mov		($Xip,&wparam(0));
1185	&mov		($Htbl,&wparam(1));
1186	&mov		($inp,&wparam(2));
1187	&mov		($len,&wparam(3));
1188
1189	&call		(&label("pic"));
1190&set_label("pic");
1191	&blindpop	($const);
1192	&lea		($const,&DWP(&label("bswap")."-".&label("pic"),$const));
1193
1194	&movdqu		($Xi,&QWP(0,$Xip));
1195	&movdqa		($T3,&QWP(0,$const));
1196	&movdqu		($Hkey,&QWP(0,$Htbl));
1197	&pshufb		($Xi,$T3);
1198
1199	&sub		($len,0x10);
1200	&jz		(&label("odd_tail"));
1201
1202	#######
1203	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
1204	#	[(H*Ii+1) + (H*Xi+1)] mod P =
1205	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
1206	#
1207	&movdqu		($T1,&QWP(0,$inp));	# Ii
1208	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1209	&pshufb		($T1,$T3);
1210	&pshufb		($Xn,$T3);
1211	&pxor		($Xi,$T1);		# Ii+Xi
1212
1213	&clmul64x64_T3	($Xhn,$Xn,$Hkey);	# H*Ii+1
1214	&movdqu		($Hkey,&QWP(16,$Htbl));	# load H^2
1215
1216	&sub		($len,0x20);
1217	&lea		($inp,&DWP(32,$inp));	# i+=2
1218	&jbe		(&label("even_tail"));
1219
1220&set_label("mod_loop");
1221	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1222	&movdqu		($Hkey,&QWP(0,$Htbl));	# load H
1223
1224	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1225	&pxor		($Xhi,$Xhn);
1226
1227	&reduction_alg5	($Xhi,$Xi);
1228
1229	#######
1230	&movdqa		($T3,&QWP(0,$const));
1231	&movdqu		($T1,&QWP(0,$inp));	# Ii
1232	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1233	&pshufb		($T1,$T3);
1234	&pshufb		($Xn,$T3);
1235	&pxor		($Xi,$T1);		# Ii+Xi
1236
1237	&clmul64x64_T3	($Xhn,$Xn,$Hkey);	# H*Ii+1
1238	&movdqu		($Hkey,&QWP(16,$Htbl));	# load H^2
1239
1240	&sub		($len,0x20);
1241	&lea		($inp,&DWP(32,$inp));
1242	&ja		(&label("mod_loop"));
1243
1244&set_label("even_tail");
1245	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1246
1247	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1248	&pxor		($Xhi,$Xhn);
1249
1250	&reduction_alg5	($Xhi,$Xi);
1251
1252	&movdqa		($T3,&QWP(0,$const));
1253	&test		($len,$len);
1254	&jnz		(&label("done"));
1255
1256	&movdqu		($Hkey,&QWP(0,$Htbl));	# load H
1257&set_label("odd_tail");
1258	&movdqu		($T1,&QWP(0,$inp));	# Ii
1259	&pshufb		($T1,$T3);
1260	&pxor		($Xi,$T1);		# Ii+Xi
1261
1262	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1263	&reduction_alg5	($Xhi,$Xi);
1264
1265	&movdqa		($T3,&QWP(0,$const));
1266&set_label("done");
1267	&pshufb		($Xi,$T3);
1268	&movdqu		(&QWP(0,$Xip),$Xi);
1269&function_end("gcm_ghash_clmul");
1270
1271}
1272
1273&set_label("bswap",64);
1274	&data_byte(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0);
1275	&data_byte(1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2);	# 0x1c2_polynomial
1276}}	# $sse2
1277
1278&set_label("rem_4bit",64);
1279	&data_word(0,0x0000<<$S,0,0x1C20<<$S,0,0x3840<<$S,0,0x2460<<$S);
1280	&data_word(0,0x7080<<$S,0,0x6CA0<<$S,0,0x48C0<<$S,0,0x54E0<<$S);
1281	&data_word(0,0xE100<<$S,0,0xFD20<<$S,0,0xD940<<$S,0,0xC560<<$S);
1282	&data_word(0,0x9180<<$S,0,0x8DA0<<$S,0,0xA9C0<<$S,0,0xB5E0<<$S);
1283&set_label("rem_8bit",64);
1284	&data_short(0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E);
1285	&data_short(0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E);
1286	&data_short(0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E);
1287	&data_short(0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E);
1288	&data_short(0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E);
1289	&data_short(0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E);
1290	&data_short(0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E);
1291	&data_short(0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E);
1292	&data_short(0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE);
1293	&data_short(0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE);
1294	&data_short(0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE);
1295	&data_short(0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE);
1296	&data_short(0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E);
1297	&data_short(0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E);
1298	&data_short(0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE);
1299	&data_short(0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE);
1300	&data_short(0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E);
1301	&data_short(0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E);
1302	&data_short(0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E);
1303	&data_short(0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E);
1304	&data_short(0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E);
1305	&data_short(0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E);
1306	&data_short(0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E);
1307	&data_short(0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E);
1308	&data_short(0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE);
1309	&data_short(0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE);
1310	&data_short(0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE);
1311	&data_short(0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE);
1312	&data_short(0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E);
1313	&data_short(0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E);
1314	&data_short(0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE);
1315	&data_short(0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE);
1316}}}	# !$x86only
1317
1318&asciz("GHASH for x86, CRYPTOGAMS by <appro\@openssl.org>");
1319&asm_finish();
1320
1321# A question was risen about choice of vanilla MMX. Or rather why wasn't
1322# SSE2 chosen instead? In addition to the fact that MMX runs on legacy
1323# CPUs such as PIII, "4-bit" MMX version was observed to provide better
1324# performance than *corresponding* SSE2 one even on contemporary CPUs.
1325# SSE2 results were provided by Peter-Michael Hager. He maintains SSE2
1326# implementation featuring full range of lookup-table sizes, but with
1327# per-invocation lookup table setup. Latter means that table size is
1328# chosen depending on how much data is to be hashed in every given call,
1329# more data - larger table. Best reported result for Core2 is ~4 cycles
1330# per processed byte out of 64KB block. This number accounts even for
1331# 64KB table setup overhead. As discussed in gcm128.c we choose to be
1332# more conservative in respect to lookup table sizes, but how do the
1333# results compare? Minimalistic "256B" MMX version delivers ~11 cycles
1334# on same platform. As also discussed in gcm128.c, next in line "8-bit
1335# Shoup's" or "4KB" method should deliver twice the performance of
1336# "256B" one, in other words not worse than ~6 cycles per byte. It
1337# should be also be noted that in SSE2 case improvement can be "super-
1338# linear," i.e. more than twice, mostly because >>8 maps to single
1339# instruction on SSE2 register. This is unlike "4-bit" case when >>4
1340# maps to same amount of instructions in both MMX and SSE2 cases.
1341# Bottom line is that switch to SSE2 is considered to be justifiable
1342# only in case we choose to implement "8-bit" method...
1343