xref: /openbsd-src/lib/libcrypto/modes/asm/ghash-x86.pl (revision 13e262cf287165162838f36e9c633b03249fc1e0)
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 between
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 �-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	($Htbl,&wparam(1));	# load Htable
413
414	&picsetup("eax");
415	&picsymbol("eax", &label("rem_4bit"), "eax");
416
417	&movz	($Zll,&BP(15,$inp));
418
419	&call	("_mmx_gmult_4bit_inner");
420
421	&mov	($inp,&wparam(0));	# load Xi
422	&emms	();
423	&mov	(&DWP(12,$inp),$Zll);
424	&mov	(&DWP(4,$inp),$Zhl);
425	&mov	(&DWP(8,$inp),$Zlh);
426	&mov	(&DWP(0,$inp),$Zhh);
427&function_end("gcm_gmult_4bit_mmx");
428
429# Streamed version performs 20% better on P4, 7% on Opteron,
430# 10% on Core2 and PIII...
431&function_begin("gcm_ghash_4bit_mmx");
432	&mov	($Zhh,&wparam(0));	# load Xi
433	&mov	($Htbl,&wparam(1));	# load Htable
434	&mov	($inp,&wparam(2));	# load in
435	&mov	($Zlh,&wparam(3));	# load len
436
437	&picsetup("eax");
438	&picsymbol("eax", &label("rem_4bit"), "eax");
439
440	&add	($Zlh,$inp);
441	&mov	(&wparam(3),$Zlh);	# len to point at the end of input
442	&stack_push(4+1);		# +1 for stack alignment
443
444	&mov	($Zll,&DWP(12,$Zhh));	# load Xi[16]
445	&mov	($Zhl,&DWP(4,$Zhh));
446	&mov	($Zlh,&DWP(8,$Zhh));
447	&mov	($Zhh,&DWP(0,$Zhh));
448	&jmp	(&label("mmx_outer_loop"));
449
450    &set_label("mmx_outer_loop",16);
451	&xor	($Zll,&DWP(12,$inp));
452	&xor	($Zhl,&DWP(4,$inp));
453	&xor	($Zlh,&DWP(8,$inp));
454	&xor	($Zhh,&DWP(0,$inp));
455	&mov	(&wparam(2),$inp);
456	&mov	(&DWP(12,"esp"),$Zll);
457	&mov	(&DWP(4,"esp"),$Zhl);
458	&mov	(&DWP(8,"esp"),$Zlh);
459	&mov	(&DWP(0,"esp"),$Zhh);
460
461	&mov	($inp,"esp");
462	&shr	($Zll,24);
463
464	&call	("_mmx_gmult_4bit_inner");
465
466	&mov	($inp,&wparam(2));
467	&lea	($inp,&DWP(16,$inp));
468	&cmp	($inp,&wparam(3));
469	&jb	(&label("mmx_outer_loop"));
470
471	&mov	($inp,&wparam(0));	# load Xi
472	&emms	();
473	&mov	(&DWP(12,$inp),$Zll);
474	&mov	(&DWP(4,$inp),$Zhl);
475	&mov	(&DWP(8,$inp),$Zlh);
476	&mov	(&DWP(0,$inp),$Zhh);
477
478	&stack_pop(4+1);
479&function_end("gcm_ghash_4bit_mmx");
480
481}} else {{	# "June" MMX version...
482		# ... has slower "April" gcm_gmult_4bit_mmx with folded
483		# loop. This is done to conserve code size...
484$S=16;		# shift factor for rem_4bit
485
486sub mmx_loop() {
487# MMX version performs 2.8 times better on P4 (see comment in non-MMX
488# routine for further details), 40% better on Opteron and Core2, 50%
489# better on PIII... In other words effort is considered to be well
490# spent...
491    my $inp = shift;
492    my $rem_4bit = shift;
493    my $cnt = $Zhh;
494    my $nhi = $Zhl;
495    my $nlo = $Zlh;
496    my $rem = $Zll;
497
498    my ($Zlo,$Zhi) = ("mm0","mm1");
499    my $tmp = "mm2";
500
501	&xor	($nlo,$nlo);	# avoid partial register stalls on PIII
502	&mov	($nhi,$Zll);
503	&mov	(&LB($nlo),&LB($nhi));
504	&mov	($cnt,14);
505	&shl	(&LB($nlo),4);
506	&and	($nhi,0xf0);
507	&movq	($Zlo,&QWP(8,$Htbl,$nlo));
508	&movq	($Zhi,&QWP(0,$Htbl,$nlo));
509	&movd	($rem,$Zlo);
510	&jmp	(&label("mmx_loop"));
511
512    &set_label("mmx_loop",16);
513	&psrlq	($Zlo,4);
514	&and	($rem,0xf);
515	&movq	($tmp,$Zhi);
516	&psrlq	($Zhi,4);
517	&pxor	($Zlo,&QWP(8,$Htbl,$nhi));
518	&mov	(&LB($nlo),&BP(0,$inp,$cnt));
519	&psllq	($tmp,60);
520	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
521	&dec	($cnt);
522	&movd	($rem,$Zlo);
523	&pxor	($Zhi,&QWP(0,$Htbl,$nhi));
524	&mov	($nhi,$nlo);
525	&pxor	($Zlo,$tmp);
526	&js	(&label("mmx_break"));
527
528	&shl	(&LB($nlo),4);
529	&and	($rem,0xf);
530	&psrlq	($Zlo,4);
531	&and	($nhi,0xf0);
532	&movq	($tmp,$Zhi);
533	&psrlq	($Zhi,4);
534	&pxor	($Zlo,&QWP(8,$Htbl,$nlo));
535	&psllq	($tmp,60);
536	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
537	&movd	($rem,$Zlo);
538	&pxor	($Zhi,&QWP(0,$Htbl,$nlo));
539	&pxor	($Zlo,$tmp);
540	&jmp	(&label("mmx_loop"));
541
542    &set_label("mmx_break",16);
543	&shl	(&LB($nlo),4);
544	&and	($rem,0xf);
545	&psrlq	($Zlo,4);
546	&and	($nhi,0xf0);
547	&movq	($tmp,$Zhi);
548	&psrlq	($Zhi,4);
549	&pxor	($Zlo,&QWP(8,$Htbl,$nlo));
550	&psllq	($tmp,60);
551	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
552	&movd	($rem,$Zlo);
553	&pxor	($Zhi,&QWP(0,$Htbl,$nlo));
554	&pxor	($Zlo,$tmp);
555
556	&psrlq	($Zlo,4);
557	&and	($rem,0xf);
558	&movq	($tmp,$Zhi);
559	&psrlq	($Zhi,4);
560	&pxor	($Zlo,&QWP(8,$Htbl,$nhi));
561	&psllq	($tmp,60);
562	&pxor	($Zhi,&QWP(0,$rem_4bit,$rem,8));
563	&movd	($rem,$Zlo);
564	&pxor	($Zhi,&QWP(0,$Htbl,$nhi));
565	&pxor	($Zlo,$tmp);
566
567	&psrlq	($Zlo,32);	# lower part of Zlo is already there
568	&movd	($Zhl,$Zhi);
569	&psrlq	($Zhi,32);
570	&movd	($Zlh,$Zlo);
571	&movd	($Zhh,$Zhi);
572
573	&bswap	($Zll);
574	&bswap	($Zhl);
575	&bswap	($Zlh);
576	&bswap	($Zhh);
577}
578
579&function_begin("gcm_gmult_4bit_mmx");
580	&mov	($inp,&wparam(0));	# load Xi
581	&mov	($Htbl,&wparam(1));	# load Htable
582
583	&picsetup("eax");
584	&picsymbol("eax", &label("rem_4bit"), "eax");
585
586	&movz	($Zll,&BP(15,$inp));
587
588	&mmx_loop($inp,"eax");
589
590	&emms	();
591	&mov	(&DWP(12,$inp),$Zll);
592	&mov	(&DWP(4,$inp),$Zhl);
593	&mov	(&DWP(8,$inp),$Zlh);
594	&mov	(&DWP(0,$inp),$Zhh);
595&function_end("gcm_gmult_4bit_mmx");
596
597######################################################################
598# Below subroutine is "528B" variant of "4-bit" GCM GHASH function
599# (see gcm128.c for details). It provides further 20-40% performance
600# improvement over above mentioned "May" version.
601
602&static_label("rem_8bit");
603
604&function_begin("gcm_ghash_4bit_mmx");
605{ my ($Zlo,$Zhi) = ("mm7","mm6");
606  my $rem_8bit = "esi";
607  my $Htbl = "ebx";
608
609    # parameter block
610    &mov	("eax",&wparam(0));		# Xi
611    &mov	("ebx",&wparam(1));		# Htable
612    &mov	("ecx",&wparam(2));		# inp
613    &mov	("edx",&wparam(3));		# len
614    &mov	("ebp","esp");			# original %esp
615
616    &picsetup($rem_8bit);
617    &picsymbol($rem_8bit, &label("rem_8bit"), $rem_8bit);
618
619    &sub	("esp",512+16+16);		# allocate stack frame...
620    &and	("esp",-64);			# ...and align it
621    &sub	("esp",16);			# place for (u8)(H[]<<4)
622
623    &add	("edx","ecx");			# pointer to the end of input
624    &mov	(&DWP(528+16+0,"esp"),"eax");	# save Xi
625    &mov	(&DWP(528+16+8,"esp"),"edx");	# save inp+len
626    &mov	(&DWP(528+16+12,"esp"),"ebp");	# save original %esp
627
628    { my @lo  = ("mm0","mm1","mm2");
629      my @hi  = ("mm3","mm4","mm5");
630      my @tmp = ("mm6","mm7");
631      my ($off1,$off2,$i) = (0,0,);
632
633      &add	($Htbl,128);			# optimize for size
634      &lea	("edi",&DWP(16+128,"esp"));
635      &lea	("ebp",&DWP(16+256+128,"esp"));
636
637      # decompose Htable (low and high parts are kept separately),
638      # generate Htable[]>>4, (u8)(Htable[]<<4), save to stack...
639      for ($i=0;$i<18;$i++) {
640
641	&mov	("edx",&DWP(16*$i+8-128,$Htbl))		if ($i<16);
642	&movq	($lo[0],&QWP(16*$i+8-128,$Htbl))	if ($i<16);
643	&psllq	($tmp[1],60)				if ($i>1);
644	&movq	($hi[0],&QWP(16*$i+0-128,$Htbl))	if ($i<16);
645	&por	($lo[2],$tmp[1])			if ($i>1);
646	&movq	(&QWP($off1-128,"edi"),$lo[1])		if ($i>0 && $i<17);
647	&psrlq	($lo[1],4)				if ($i>0 && $i<17);
648	&movq	(&QWP($off1,"edi"),$hi[1])		if ($i>0 && $i<17);
649	&movq	($tmp[0],$hi[1])			if ($i>0 && $i<17);
650	&movq	(&QWP($off2-128,"ebp"),$lo[2])		if ($i>1);
651	&psrlq	($hi[1],4)				if ($i>0 && $i<17);
652	&movq	(&QWP($off2,"ebp"),$hi[2])		if ($i>1);
653	&shl	("edx",4)				if ($i<16);
654	&mov	(&BP($i,"esp"),&LB("edx"))		if ($i<16);
655
656	unshift	(@lo,pop(@lo));			# "rotate" registers
657	unshift	(@hi,pop(@hi));
658	unshift	(@tmp,pop(@tmp));
659	$off1 += 8	if ($i>0);
660	$off2 += 8	if ($i>1);
661      }
662    }
663
664    &movq	($Zhi,&QWP(0,"eax"));
665    &mov	("ebx",&DWP(8,"eax"));
666    &mov	("edx",&DWP(12,"eax"));		# load Xi
667
668&set_label("outer",16);
669  { my $nlo = "eax";
670    my $dat = "edx";
671    my @nhi = ("edi","ebp");
672    my @rem = ("ebx","ecx");
673    my @red = ("mm0","mm1","mm2");
674    my $tmp = "mm3";
675
676    &xor	($dat,&DWP(12,"ecx"));		# merge input data
677    &xor	("ebx",&DWP(8,"ecx"));
678    &pxor	($Zhi,&QWP(0,"ecx"));
679    &lea	("ecx",&DWP(16,"ecx"));		# inp+=16
680    #&mov	(&DWP(528+12,"esp"),$dat);	# save inp^Xi
681    &mov	(&DWP(528+8,"esp"),"ebx");
682    &movq	(&QWP(528+0,"esp"),$Zhi);
683    &mov	(&DWP(528+16+4,"esp"),"ecx");	# save inp
684
685    &xor	($nlo,$nlo);
686    &rol	($dat,8);
687    &mov	(&LB($nlo),&LB($dat));
688    &mov	($nhi[1],$nlo);
689    &and	(&LB($nlo),0x0f);
690    &shr	($nhi[1],4);
691    &pxor	($red[0],$red[0]);
692    &rol	($dat,8);			# next byte
693    &pxor	($red[1],$red[1]);
694    &pxor	($red[2],$red[2]);
695
696    # Just like in "May" version modulo-schedule for critical path in
697    # 'Z.hi ^= rem_8bit[Z.lo&0xff^((u8)H[nhi]<<4)]<<48'. Final 'pxor'
698    # is scheduled so late that rem_8bit[] has to be shifted *right*
699    # by 16, which is why last argument to pinsrw is 2, which
700    # corresponds to <<32=<<48>>16...
701    for ($j=11,$i=0;$i<15;$i++) {
702
703      if ($i>0) {
704	&pxor	($Zlo,&QWP(16,"esp",$nlo,8));		# Z^=H[nlo]
705	&rol	($dat,8);				# next byte
706	&pxor	($Zhi,&QWP(16+128,"esp",$nlo,8));
707
708	&pxor	($Zlo,$tmp);
709	&pxor	($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
710	&xor	(&LB($rem[1]),&BP(0,"esp",$nhi[0]));	# rem^(H[nhi]<<4)
711      } else {
712	&movq	($Zlo,&QWP(16,"esp",$nlo,8));
713	&movq	($Zhi,&QWP(16+128,"esp",$nlo,8));
714      }
715
716	&mov	(&LB($nlo),&LB($dat));
717	&mov	($dat,&DWP(528+$j,"esp"))		if (--$j%4==0 && $j>=0);
718
719	&movd	($rem[0],$Zlo);
720	&movz	($rem[1],&LB($rem[1]))			if ($i>0);
721	&psrlq	($Zlo,8);				# Z>>=8
722
723	&movq	($tmp,$Zhi);
724	&mov	($nhi[0],$nlo);
725	&psrlq	($Zhi,8);
726
727	&pxor	($Zlo,&QWP(16+256+0,"esp",$nhi[1],8));	# Z^=H[nhi]>>4
728	&and	(&LB($nlo),0x0f);
729	&psllq	($tmp,56);
730
731	&pxor	($Zhi,$red[1])				if ($i>1);
732	&shr	($nhi[0],4);
733	&pinsrw	($red[0],&WP(0,$rem_8bit,$rem[1],2),2)	if ($i>0);
734
735	unshift	(@red,pop(@red));			# "rotate" registers
736	unshift	(@rem,pop(@rem));
737	unshift	(@nhi,pop(@nhi));
738    }
739
740    &pxor	($Zlo,&QWP(16,"esp",$nlo,8));		# Z^=H[nlo]
741    &pxor	($Zhi,&QWP(16+128,"esp",$nlo,8));
742    &xor	(&LB($rem[1]),&BP(0,"esp",$nhi[0]));	# rem^(H[nhi]<<4)
743
744    &pxor	($Zlo,$tmp);
745    &pxor	($Zhi,&QWP(16+256+128,"esp",$nhi[0],8));
746    &movz	($rem[1],&LB($rem[1]));
747
748    &pxor	($red[2],$red[2]);			# clear 2nd word
749    &psllq	($red[1],4);
750
751    &movd	($rem[0],$Zlo);
752    &psrlq	($Zlo,4);				# Z>>=4
753
754    &movq	($tmp,$Zhi);
755    &psrlq	($Zhi,4);
756    &shl	($rem[0],4);				# rem<<4
757
758    &pxor	($Zlo,&QWP(16,"esp",$nhi[1],8));	# Z^=H[nhi]
759    &psllq	($tmp,60);
760    &movz	($rem[0],&LB($rem[0]));
761
762    &pxor	($Zlo,$tmp);
763    &pxor	($Zhi,&QWP(16+128,"esp",$nhi[1],8));
764
765    &pinsrw	($red[0],&WP(0,$rem_8bit,$rem[1],2),2);
766    &pxor	($Zhi,$red[1]);
767
768    &movd	($dat,$Zlo);
769    &pinsrw	($red[2],&WP(0,$rem_8bit,$rem[0],2),3);	# last is <<48
770
771    &psllq	($red[0],12);				# correct by <<16>>4
772    &pxor	($Zhi,$red[0]);
773    &psrlq	($Zlo,32);
774    &pxor	($Zhi,$red[2]);
775
776    &mov	("ecx",&DWP(528+16+4,"esp"));	# restore inp
777    &movd	("ebx",$Zlo);
778    &movq	($tmp,$Zhi);			# 01234567
779    &psllw	($Zhi,8);			# 1.3.5.7.
780    &psrlw	($tmp,8);			# .0.2.4.6
781    &por	($Zhi,$tmp);			# 10325476
782    &bswap	($dat);
783    &pshufw	($Zhi,$Zhi,0b00011011);		# 76543210
784    &bswap	("ebx");
785
786    &cmp	("ecx",&DWP(528+16+8,"esp"));	# are we done?
787    &jne	(&label("outer"));
788  }
789
790    &mov	("eax",&DWP(528+16+0,"esp"));	# restore Xi
791    &mov	(&DWP(12,"eax"),"edx");
792    &mov	(&DWP(8,"eax"),"ebx");
793    &movq	(&QWP(0,"eax"),$Zhi);
794
795    &mov	("esp",&DWP(528+16+12,"esp"));	# restore original %esp
796    &emms	();
797}
798&function_end("gcm_ghash_4bit_mmx");
799}}
800
801if ($sse2) {{
802######################################################################
803# PCLMULQDQ version.
804
805$Xip="eax";
806$Htbl="edx";
807$const="ecx";
808$inp="esi";
809$len="ebx";
810
811($Xi,$Xhi)=("xmm0","xmm1");	$Hkey="xmm2";
812($T1,$T2,$T3)=("xmm3","xmm4","xmm5");
813($Xn,$Xhn)=("xmm6","xmm7");
814
815&static_label("bswap");
816
817sub clmul64x64_T2 {	# minimal "register" pressure
818my ($Xhi,$Xi,$Hkey)=@_;
819
820	&movdqa		($Xhi,$Xi);		#
821	&pshufd		($T1,$Xi,0b01001110);
822	&pshufd		($T2,$Hkey,0b01001110);
823	&pxor		($T1,$Xi);		#
824	&pxor		($T2,$Hkey);
825
826	&pclmulqdq	($Xi,$Hkey,0x00);	#######
827	&pclmulqdq	($Xhi,$Hkey,0x11);	#######
828	&pclmulqdq	($T1,$T2,0x00);		#######
829	&xorps		($T1,$Xi);		#
830	&xorps		($T1,$Xhi);		#
831
832	&movdqa		($T2,$T1);		#
833	&psrldq		($T1,8);
834	&pslldq		($T2,8);		#
835	&pxor		($Xhi,$T1);
836	&pxor		($Xi,$T2);		#
837}
838
839sub clmul64x64_T3 {
840# Even though this subroutine offers visually better ILP, it
841# was empirically found to be a tad slower than above version.
842# At least in gcm_ghash_clmul context. But it's just as well,
843# because loop modulo-scheduling is possible only thanks to
844# minimized "register" pressure...
845my ($Xhi,$Xi,$Hkey)=@_;
846
847	&movdqa		($T1,$Xi);		#
848	&movdqa		($Xhi,$Xi);
849	&pclmulqdq	($Xi,$Hkey,0x00);	#######
850	&pclmulqdq	($Xhi,$Hkey,0x11);	#######
851	&pshufd		($T2,$T1,0b01001110);	#
852	&pshufd		($T3,$Hkey,0b01001110);
853	&pxor		($T2,$T1);		#
854	&pxor		($T3,$Hkey);
855	&pclmulqdq	($T2,$T3,0x00);		#######
856	&pxor		($T2,$Xi);		#
857	&pxor		($T2,$Xhi);		#
858
859	&movdqa		($T3,$T2);		#
860	&psrldq		($T2,8);
861	&pslldq		($T3,8);		#
862	&pxor		($Xhi,$T2);
863	&pxor		($Xi,$T3);		#
864}
865
866if (1) {		# Algorithm 9 with <<1 twist.
867			# Reduction is shorter and uses only two
868			# temporary registers, which makes it better
869			# candidate for interleaving with 64x64
870			# multiplication. Pre-modulo-scheduled loop
871			# was found to be ~20% faster than Algorithm 5
872			# below. Algorithm 9 was therefore chosen for
873			# further optimization...
874
875sub reduction_alg9 {	# 17/13 times faster than Intel version
876my ($Xhi,$Xi) = @_;
877
878	# 1st phase
879	&movdqa		($T1,$Xi);		#
880	&psllq		($Xi,1);
881	&pxor		($Xi,$T1);		#
882	&psllq		($Xi,5);		#
883	&pxor		($Xi,$T1);		#
884	&psllq		($Xi,57);		#
885	&movdqa		($T2,$Xi);		#
886	&pslldq		($Xi,8);
887	&psrldq		($T2,8);		#
888	&pxor		($Xi,$T1);
889	&pxor		($Xhi,$T2);		#
890
891	# 2nd phase
892	&movdqa		($T2,$Xi);
893	&psrlq		($Xi,5);
894	&pxor		($Xi,$T2);		#
895	&psrlq		($Xi,1);		#
896	&pxor		($Xi,$T2);		#
897	&pxor		($T2,$Xhi);
898	&psrlq		($Xi,1);		#
899	&pxor		($Xi,$T2);		#
900}
901
902&function_begin_B("gcm_init_clmul");
903	&mov		($Htbl,&wparam(0));
904	&mov		($Xip,&wparam(1));
905
906	&picsetup($const);
907	&picsymbol($const, &label("bswap"), $const);
908
909	&movdqu		($Hkey,&QWP(0,$Xip));
910	&pshufd		($Hkey,$Hkey,0b01001110);# dword swap
911
912	# <<1 twist
913	&pshufd		($T2,$Hkey,0b11111111);	# broadcast uppermost dword
914	&movdqa		($T1,$Hkey);
915	&psllq		($Hkey,1);
916	&pxor		($T3,$T3);		#
917	&psrlq		($T1,63);
918	&pcmpgtd	($T3,$T2);		# broadcast carry bit
919	&pslldq		($T1,8);
920	&por		($Hkey,$T1);		# H<<=1
921
922	# magic reduction
923	&pand		($T3,&QWP(16,$const));	# 0x1c2_polynomial
924	&pxor		($Hkey,$T3);		# if(carry) H^=0x1c2_polynomial
925
926	# calculate H^2
927	&movdqa		($Xi,$Hkey);
928	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
929	&reduction_alg9	($Xhi,$Xi);
930
931	&movdqu		(&QWP(0,$Htbl),$Hkey);	# save H
932	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
933
934	&ret		();
935&function_end_B("gcm_init_clmul");
936
937&function_begin_B("gcm_gmult_clmul");
938	&mov		($Xip,&wparam(0));
939	&mov		($Htbl,&wparam(1));
940
941	&picsetup($const);
942	&picsymbol($const, &label("bswap"), $const);
943
944	&movdqu		($Xi,&QWP(0,$Xip));
945	&movdqa		($T3,&QWP(0,$const));
946	&movups		($Hkey,&QWP(0,$Htbl));
947	&pshufb		($Xi,$T3);
948
949	&clmul64x64_T2	($Xhi,$Xi,$Hkey);
950	&reduction_alg9	($Xhi,$Xi);
951
952	&pshufb		($Xi,$T3);
953	&movdqu		(&QWP(0,$Xip),$Xi);
954
955	&ret	();
956&function_end_B("gcm_gmult_clmul");
957
958&function_begin("gcm_ghash_clmul");
959	&mov		($Xip,&wparam(0));
960	&mov		($Htbl,&wparam(1));
961	&mov		($inp,&wparam(2));
962	&mov		($len,&wparam(3));
963
964	&picsetup($const);
965	&picsymbol($const, &label("bswap"), $const);
966
967	&movdqu		($Xi,&QWP(0,$Xip));
968	&movdqa		($T3,&QWP(0,$const));
969	&movdqu		($Hkey,&QWP(0,$Htbl));
970	&pshufb		($Xi,$T3);
971
972	&sub		($len,0x10);
973	&jz		(&label("odd_tail"));
974
975	#######
976	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
977	#	[(H*Ii+1) + (H*Xi+1)] mod P =
978	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
979	#
980	&movdqu		($T1,&QWP(0,$inp));	# Ii
981	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
982	&pshufb		($T1,$T3);
983	&pshufb		($Xn,$T3);
984	&pxor		($Xi,$T1);		# Ii+Xi
985
986	&clmul64x64_T2	($Xhn,$Xn,$Hkey);	# H*Ii+1
987	&movups		($Hkey,&QWP(16,$Htbl));	# load H^2
988
989	&lea		($inp,&DWP(32,$inp));	# i+=2
990	&sub		($len,0x20);
991	&jbe		(&label("even_tail"));
992
993&set_label("mod_loop");
994	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
995	&movdqu		($T1,&QWP(0,$inp));	# Ii
996	&movups		($Hkey,&QWP(0,$Htbl));	# load H
997
998	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
999	&pxor		($Xhi,$Xhn);
1000
1001	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1002	&pshufb		($T1,$T3);
1003	&pshufb		($Xn,$T3);
1004
1005	&movdqa		($T3,$Xn);		#&clmul64x64_TX	($Xhn,$Xn,$Hkey); H*Ii+1
1006	&movdqa		($Xhn,$Xn);
1007	 &pxor		($Xhi,$T1);		# "Ii+Xi", consume early
1008
1009	  &movdqa	($T1,$Xi);		#&reduction_alg9($Xhi,$Xi); 1st phase
1010	  &psllq	($Xi,1);
1011	  &pxor		($Xi,$T1);		#
1012	  &psllq	($Xi,5);		#
1013	  &pxor		($Xi,$T1);		#
1014	&pclmulqdq	($Xn,$Hkey,0x00);	#######
1015	  &psllq	($Xi,57);		#
1016	  &movdqa	($T2,$Xi);		#
1017	  &pslldq	($Xi,8);
1018	  &psrldq	($T2,8);		#
1019	  &pxor		($Xi,$T1);
1020	&pshufd		($T1,$T3,0b01001110);
1021	  &pxor		($Xhi,$T2);		#
1022	&pxor		($T1,$T3);
1023	&pshufd		($T3,$Hkey,0b01001110);
1024	&pxor		($T3,$Hkey);		#
1025
1026	&pclmulqdq	($Xhn,$Hkey,0x11);	#######
1027	  &movdqa	($T2,$Xi);		# 2nd phase
1028	  &psrlq	($Xi,5);
1029	  &pxor		($Xi,$T2);		#
1030	  &psrlq	($Xi,1);		#
1031	  &pxor		($Xi,$T2);		#
1032	  &pxor		($T2,$Xhi);
1033	  &psrlq	($Xi,1);		#
1034	  &pxor		($Xi,$T2);		#
1035
1036	&pclmulqdq	($T1,$T3,0x00);		#######
1037	&movups		($Hkey,&QWP(16,$Htbl));	# load H^2
1038	&xorps		($T1,$Xn);		#
1039	&xorps		($T1,$Xhn);		#
1040
1041	&movdqa		($T3,$T1);		#
1042	&psrldq		($T1,8);
1043	&pslldq		($T3,8);		#
1044	&pxor		($Xhn,$T1);
1045	&pxor		($Xn,$T3);		#
1046	&movdqa		($T3,&QWP(0,$const));
1047
1048	&lea		($inp,&DWP(32,$inp));
1049	&sub		($len,0x20);
1050	&ja		(&label("mod_loop"));
1051
1052&set_label("even_tail");
1053	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1054
1055	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1056	&pxor		($Xhi,$Xhn);
1057
1058	&reduction_alg9	($Xhi,$Xi);
1059
1060	&test		($len,$len);
1061	&jnz		(&label("done"));
1062
1063	&movups		($Hkey,&QWP(0,$Htbl));	# load H
1064&set_label("odd_tail");
1065	&movdqu		($T1,&QWP(0,$inp));	# Ii
1066	&pshufb		($T1,$T3);
1067	&pxor		($Xi,$T1);		# Ii+Xi
1068
1069	&clmul64x64_T2	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1070	&reduction_alg9	($Xhi,$Xi);
1071
1072&set_label("done");
1073	&pshufb		($Xi,$T3);
1074	&movdqu		(&QWP(0,$Xip),$Xi);
1075&function_end("gcm_ghash_clmul");
1076
1077} else {		# Algorithm 5. Kept for reference purposes.
1078
1079sub reduction_alg5 {	# 19/16 times faster than Intel version
1080my ($Xhi,$Xi)=@_;
1081
1082	# <<1
1083	&movdqa		($T1,$Xi);		#
1084	&movdqa		($T2,$Xhi);
1085	&pslld		($Xi,1);
1086	&pslld		($Xhi,1);		#
1087	&psrld		($T1,31);
1088	&psrld		($T2,31);		#
1089	&movdqa		($T3,$T1);
1090	&pslldq		($T1,4);
1091	&psrldq		($T3,12);		#
1092	&pslldq		($T2,4);
1093	&por		($Xhi,$T3);		#
1094	&por		($Xi,$T1);
1095	&por		($Xhi,$T2);		#
1096
1097	# 1st phase
1098	&movdqa		($T1,$Xi);
1099	&movdqa		($T2,$Xi);
1100	&movdqa		($T3,$Xi);		#
1101	&pslld		($T1,31);
1102	&pslld		($T2,30);
1103	&pslld		($Xi,25);		#
1104	&pxor		($T1,$T2);
1105	&pxor		($T1,$Xi);		#
1106	&movdqa		($T2,$T1);		#
1107	&pslldq		($T1,12);
1108	&psrldq		($T2,4);		#
1109	&pxor		($T3,$T1);
1110
1111	# 2nd phase
1112	&pxor		($Xhi,$T3);		#
1113	&movdqa		($Xi,$T3);
1114	&movdqa		($T1,$T3);
1115	&psrld		($Xi,1);		#
1116	&psrld		($T1,2);
1117	&psrld		($T3,7);		#
1118	&pxor		($Xi,$T1);
1119	&pxor		($Xhi,$T2);
1120	&pxor		($Xi,$T3);		#
1121	&pxor		($Xi,$Xhi);		#
1122}
1123
1124&function_begin_B("gcm_init_clmul");
1125	&mov		($Htbl,&wparam(0));
1126	&mov		($Xip,&wparam(1));
1127
1128	&picsetup($const);
1129	&picsymbol($const, &label("bswap"), $const);
1130
1131	&movdqu		($Hkey,&QWP(0,$Xip));
1132	&pshufd		($Hkey,$Hkey,0b01001110);# dword swap
1133
1134	# calculate H^2
1135	&movdqa		($Xi,$Hkey);
1136	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1137	&reduction_alg5	($Xhi,$Xi);
1138
1139	&movdqu		(&QWP(0,$Htbl),$Hkey);	# save H
1140	&movdqu		(&QWP(16,$Htbl),$Xi);	# save H^2
1141
1142	&ret		();
1143&function_end_B("gcm_init_clmul");
1144
1145&function_begin_B("gcm_gmult_clmul");
1146	&mov		($Xip,&wparam(0));
1147	&mov		($Htbl,&wparam(1));
1148
1149	&picsetup($const);
1150	&picsymbol($const, &label("bswap"), $const);
1151
1152	&movdqu		($Xi,&QWP(0,$Xip));
1153	&movdqa		($Xn,&QWP(0,$const));
1154	&movdqu		($Hkey,&QWP(0,$Htbl));
1155	&pshufb		($Xi,$Xn);
1156
1157	&clmul64x64_T3	($Xhi,$Xi,$Hkey);
1158	&reduction_alg5	($Xhi,$Xi);
1159
1160	&pshufb		($Xi,$Xn);
1161	&movdqu		(&QWP(0,$Xip),$Xi);
1162
1163	&ret	();
1164&function_end_B("gcm_gmult_clmul");
1165
1166&function_begin("gcm_ghash_clmul");
1167	&mov		($Xip,&wparam(0));
1168	&mov		($Htbl,&wparam(1));
1169	&mov		($inp,&wparam(2));
1170	&mov		($len,&wparam(3));
1171
1172	&picsetup($const);
1173	&picsymbol($const, &label("bswap"), $const);
1174
1175	&movdqu		($Xi,&QWP(0,$Xip));
1176	&movdqa		($T3,&QWP(0,$const));
1177	&movdqu		($Hkey,&QWP(0,$Htbl));
1178	&pshufb		($Xi,$T3);
1179
1180	&sub		($len,0x10);
1181	&jz		(&label("odd_tail"));
1182
1183	#######
1184	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
1185	#	[(H*Ii+1) + (H*Xi+1)] mod P =
1186	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
1187	#
1188	&movdqu		($T1,&QWP(0,$inp));	# Ii
1189	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1190	&pshufb		($T1,$T3);
1191	&pshufb		($Xn,$T3);
1192	&pxor		($Xi,$T1);		# Ii+Xi
1193
1194	&clmul64x64_T3	($Xhn,$Xn,$Hkey);	# H*Ii+1
1195	&movdqu		($Hkey,&QWP(16,$Htbl));	# load H^2
1196
1197	&sub		($len,0x20);
1198	&lea		($inp,&DWP(32,$inp));	# i+=2
1199	&jbe		(&label("even_tail"));
1200
1201&set_label("mod_loop");
1202	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1203	&movdqu		($Hkey,&QWP(0,$Htbl));	# load H
1204
1205	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1206	&pxor		($Xhi,$Xhn);
1207
1208	&reduction_alg5	($Xhi,$Xi);
1209
1210	#######
1211	&movdqa		($T3,&QWP(0,$const));
1212	&movdqu		($T1,&QWP(0,$inp));	# Ii
1213	&movdqu		($Xn,&QWP(16,$inp));	# Ii+1
1214	&pshufb		($T1,$T3);
1215	&pshufb		($Xn,$T3);
1216	&pxor		($Xi,$T1);		# Ii+Xi
1217
1218	&clmul64x64_T3	($Xhn,$Xn,$Hkey);	# H*Ii+1
1219	&movdqu		($Hkey,&QWP(16,$Htbl));	# load H^2
1220
1221	&sub		($len,0x20);
1222	&lea		($inp,&DWP(32,$inp));
1223	&ja		(&label("mod_loop"));
1224
1225&set_label("even_tail");
1226	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H^2*(Ii+Xi)
1227
1228	&pxor		($Xi,$Xn);		# (H*Ii+1) + H^2*(Ii+Xi)
1229	&pxor		($Xhi,$Xhn);
1230
1231	&reduction_alg5	($Xhi,$Xi);
1232
1233	&movdqa		($T3,&QWP(0,$const));
1234	&test		($len,$len);
1235	&jnz		(&label("done"));
1236
1237	&movdqu		($Hkey,&QWP(0,$Htbl));	# load H
1238&set_label("odd_tail");
1239	&movdqu		($T1,&QWP(0,$inp));	# Ii
1240	&pshufb		($T1,$T3);
1241	&pxor		($Xi,$T1);		# Ii+Xi
1242
1243	&clmul64x64_T3	($Xhi,$Xi,$Hkey);	# H*(Ii+Xi)
1244	&reduction_alg5	($Xhi,$Xi);
1245
1246	&movdqa		($T3,&QWP(0,$const));
1247&set_label("done");
1248	&pshufb		($Xi,$T3);
1249	&movdqu		(&QWP(0,$Xip),$Xi);
1250&function_end("gcm_ghash_clmul");
1251
1252}
1253
1254	&rodataseg();
1255&set_label("bswap",64);
1256	&data_byte(15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0);
1257	&data_byte(1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2);	# 0x1c2_polynomial
1258	&previous();
1259}}	# $sse2
1260
1261	&rodataseg();
1262&set_label("rem_4bit",64);
1263	&data_word(0,0x0000<<$S,0,0x1C20<<$S,0,0x3840<<$S,0,0x2460<<$S);
1264	&data_word(0,0x7080<<$S,0,0x6CA0<<$S,0,0x48C0<<$S,0,0x54E0<<$S);
1265	&data_word(0,0xE100<<$S,0,0xFD20<<$S,0,0xD940<<$S,0,0xC560<<$S);
1266	&data_word(0,0x9180<<$S,0,0x8DA0<<$S,0,0xA9C0<<$S,0,0xB5E0<<$S);
1267&set_label("rem_8bit",64);
1268	&data_short(0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E);
1269	&data_short(0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E);
1270	&data_short(0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E);
1271	&data_short(0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E);
1272	&data_short(0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E);
1273	&data_short(0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E);
1274	&data_short(0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E);
1275	&data_short(0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E);
1276	&data_short(0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE);
1277	&data_short(0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE);
1278	&data_short(0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE);
1279	&data_short(0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE);
1280	&data_short(0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E);
1281	&data_short(0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E);
1282	&data_short(0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE);
1283	&data_short(0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE);
1284	&data_short(0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E);
1285	&data_short(0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E);
1286	&data_short(0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E);
1287	&data_short(0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E);
1288	&data_short(0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E);
1289	&data_short(0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E);
1290	&data_short(0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E);
1291	&data_short(0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E);
1292	&data_short(0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE);
1293	&data_short(0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE);
1294	&data_short(0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE);
1295	&data_short(0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE);
1296	&data_short(0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E);
1297	&data_short(0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E);
1298	&data_short(0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE);
1299	&data_short(0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE);
1300	&previous();
1301}}}	# !$x86only
1302
1303&asm_finish();
1304
1305# A question was risen about choice of vanilla MMX. Or rather why wasn't
1306# SSE2 chosen instead? In addition to the fact that MMX runs on legacy
1307# CPUs such as PIII, "4-bit" MMX version was observed to provide better
1308# performance than *corresponding* SSE2 one even on contemporary CPUs.
1309# SSE2 results were provided by Peter-Michael Hager. He maintains SSE2
1310# implementation featuring full range of lookup-table sizes, but with
1311# per-invocation lookup table setup. Latter means that table size is
1312# chosen depending on how much data is to be hashed in every given call,
1313# more data - larger table. Best reported result for Core2 is ~4 cycles
1314# per processed byte out of 64KB block. This number accounts even for
1315# 64KB table setup overhead. As discussed in gcm128.c we choose to be
1316# more conservative in respect to lookup table sizes, but how do the
1317# results compare? Minimalistic "256B" MMX version delivers ~11 cycles
1318# on same platform. As also discussed in gcm128.c, next in line "8-bit
1319# Shoup's" or "4KB" method should deliver twice the performance of
1320# "256B" one, in other words not worse than ~6 cycles per byte. It
1321# should be also be noted that in SSE2 case improvement can be "super-
1322# linear," i.e. more than twice, mostly because >>8 maps to single
1323# instruction on SSE2 register. This is unlike "4-bit" case when >>4
1324# maps to same amount of instructions in both MMX and SSE2 cases.
1325# Bottom line is that switch to SSE2 is considered to be justifiable
1326# only in case we choose to implement "8-bit" method...
1327