xref: /freebsd/crypto/openssl/crypto/sha/asm/keccak1600-mmx.pl (revision b9f654b163bce26de79705e77b872427c9f2afa1)
1#!/usr/bin/env perl
2# Copyright 2017 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8#
9# ====================================================================
10# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
11# project. The module is, however, dual licensed under OpenSSL and
12# CRYPTOGAMS licenses depending on where you obtain it. For further
13# details see http://www.openssl.org/~appro/cryptogams/.
14# ====================================================================
15#
16# Keccak-1600 for x86 MMX.
17#
18# June 2017.
19#
20# Below code is KECCAK_2X implementation (see sha/keccak1600.c) with
21# C[5] held in register bank and D[5] offloaded to memory. Though
22# instead of actually unrolling the loop pair-wise I simply flip
23# pointers to T[][] and A[][] and the end of round. Since number of
24# rounds is even, last round writes to A[][] and everything works out.
25# It's argued that MMX is the only code path meaningful to implement
26# for x86. This is because non-MMX-capable processors is an extinct
27# breed, and they as well can lurk executing compiler-generated code.
28# For reference gcc-5.x-generated KECCAK_2X code takes 89 cycles per
29# processed byte on Pentium. Which is fair result. But older compilers
30# produce worse code. On the other hand one can wonder why not 128-bit
31# SSE2? Well, SSE2 won't provide double improvement, rather far from
32# that, if any at all on some processors, because it will take extra
33# permutations and inter-bank data trasfers. Besides, contemporary
34# CPUs are better off executing 64-bit code, and it makes lesser sense
35# to invest into fancy 32-bit code. And the decision doesn't seem to
36# be inadequate, if one compares below results to "64-bit platforms in
37# 32-bit mode" SIMD data points available at
38# http://keccak.noekeon.org/sw_performance.html.
39#
40########################################################################
41# Numbers are cycles per processed byte out of large message.
42#
43#			r=1088(i)
44#
45# PIII			30/+150%
46# Pentium M		27/+150%
47# P4			40/+85%
48# Core 2		19/+170%
49# Sandy Bridge(ii)	18/+140%
50# Atom			33/+180%
51# Silvermont(ii)	30/+180%
52# VIA Nano(ii)		43/+60%
53# Sledgehammer(ii)(iii)	24/+130%
54#
55# (i)	Corresponds to SHA3-256. Numbers after slash are improvement
56#	coefficients over KECCAK_2X [with bit interleave and lane
57#	complementing] position-independent *scalar* code generated
58#	by gcc-5.x. It's not exactly fair comparison, but it's a
59#	datapoint...
60# (ii)	64-bit processor executing 32-bit code.
61# (iii)	Result is considered to be representative even for older AMD
62#	processors.
63
64$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
65push(@INC,"${dir}","${dir}../../perlasm");
66require "x86asm.pl";
67
68$output=pop;
69open STDOUT,">$output";
70
71&asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
72
73my @C = map("mm$_",(0..4));
74my @T = map("mm$_",(5..7));
75my @A = map([ 8*$_-100, 8*($_+1)-100, 8*($_+2)-100,
76              8*($_+3)-100, 8*($_+4)-100 ], (0,5,10,15,20));
77my @D = map(8*$_+4, (0..4));
78my @rhotates = ([  0,  1, 62, 28, 27 ],
79                [ 36, 44,  6, 55, 20 ],
80                [  3, 10, 43, 25, 39 ],
81                [ 41, 45, 15, 21,  8 ],
82                [ 18,  2, 61, 56, 14 ]);
83
84&static_label("iotas");
85
86&function_begin_B("_KeccakF1600");
87	&movq	(@C[0],&QWP($A[4][0],"esi"));
88	&movq	(@C[1],&QWP($A[4][1],"esi"));
89	&movq	(@C[2],&QWP($A[4][2],"esi"));
90	&movq	(@C[3],&QWP($A[4][3],"esi"));
91	&movq	(@C[4],&QWP($A[4][4],"esi"));
92
93	&mov	("ecx",24);			# loop counter
94	&jmp	(&label("loop"));
95
96    &set_label("loop",16);
97	######################################### Theta
98	&pxor	(@C[0],&QWP($A[0][0],"esi"));
99	&pxor	(@C[1],&QWP($A[0][1],"esi"));
100	&pxor	(@C[2],&QWP($A[0][2],"esi"));
101	&pxor	(@C[3],&QWP($A[0][3],"esi"));
102	&pxor	(@C[4],&QWP($A[0][4],"esi"));
103
104	&pxor	(@C[0],&QWP($A[1][0],"esi"));
105	&pxor	(@C[1],&QWP($A[1][1],"esi"));
106	&pxor	(@C[2],&QWP($A[1][2],"esi"));
107	&pxor	(@C[3],&QWP($A[1][3],"esi"));
108	&pxor	(@C[4],&QWP($A[1][4],"esi"));
109
110	&pxor	(@C[0],&QWP($A[2][0],"esi"));
111	&pxor	(@C[1],&QWP($A[2][1],"esi"));
112	&pxor	(@C[2],&QWP($A[2][2],"esi"));
113	&pxor	(@C[3],&QWP($A[2][3],"esi"));
114	&pxor	(@C[4],&QWP($A[2][4],"esi"));
115
116	&pxor	(@C[2],&QWP($A[3][2],"esi"));
117	&pxor	(@C[0],&QWP($A[3][0],"esi"));
118	&pxor	(@C[1],&QWP($A[3][1],"esi"));
119	&pxor	(@C[3],&QWP($A[3][3],"esi"));
120	 &movq	(@T[0],@C[2]);
121	&pxor	(@C[4],&QWP($A[3][4],"esi"));
122
123	 &movq	(@T[2],@C[2]);
124	 &psrlq	(@T[0],63);
125	&movq	(@T[1],@C[0]);
126	 &psllq	(@T[2],1);
127	 &pxor	(@T[0],@C[0]);
128	&psrlq	(@C[0],63);
129	 &pxor	(@T[0],@T[2]);
130	&psllq	(@T[1],1);
131	 &movq	(@T[2],@C[1]);
132	 &movq	(&QWP(@D[1],"esp"),@T[0]);	# D[1] = E[0] = ROL64(C[2], 1) ^ C[0];
133
134	&pxor	(@T[1],@C[0]);
135	 &psrlq	(@T[2],63);
136	&pxor	(@T[1],@C[3]);
137	 &movq	(@C[0],@C[1]);
138	&movq	(&QWP(@D[4],"esp"),@T[1]);	# D[4] = E[1] = ROL64(C[0], 1) ^ C[3];
139
140	 &psllq	(@C[0],1);
141	 &pxor	(@T[2],@C[4]);
142	 &pxor	(@C[0],@T[2]);
143
144	&movq	(@T[2],@C[3]);
145	&psrlq	(@C[3],63);
146	 &movq	(&QWP(@D[0],"esp"),@C[0]);	# D[0] = C[0] = ROL64(C[1], 1) ^ C[4];
147	&psllq	(@T[2],1);
148	 &movq	(@T[0],@C[4]);
149	 &psrlq	(@C[4],63);
150	&pxor	(@C[1],@C[3]);
151	 &psllq	(@T[0],1);
152	&pxor	(@C[1],@T[2]);
153	 &pxor	(@C[2],@C[4]);
154	&movq	(&QWP(@D[2],"esp"),@C[1]);	# D[2] = C[1] = ROL64(C[3], 1) ^ C[1];
155	 &pxor	(@C[2],@T[0]);
156
157	######################################### first Rho(0) is special
158	&movq	(@C[3],&QWP($A[3][3],"esi"));
159	 &movq	(&QWP(@D[3],"esp"),@C[2]);	# D[3] = C[2] = ROL64(C[4], 1) ^ C[2];
160	&pxor	(@C[3],@C[2]);
161	 &movq	(@C[4],&QWP($A[4][4],"esi"));
162	&movq	(@T[2],@C[3]);
163	&psrlq	(@C[3],64-$rhotates[3][3]);
164	 &pxor	(@C[4],@T[1]);
165	&psllq	(@T[2],$rhotates[3][3]);
166	 &movq	(@T[1],@C[4]);
167	 &psrlq	(@C[4],64-$rhotates[4][4]);
168	&por	(@C[3],@T[2]);		# C[3] = ROL64(A[3][3] ^ C[2], rhotates[3][3]);   /* D[3] */
169	 &psllq	(@T[1],$rhotates[4][4]);
170
171	&movq	(@C[2],&QWP($A[2][2],"esi"));
172	 &por	(@C[4],@T[1]);		# C[4] = ROL64(A[4][4] ^ E[1], rhotates[4][4]);   /* D[4] */
173	&pxor	(@C[2],@C[1]);
174	 &movq	(@C[1],&QWP($A[1][1],"esi"));
175	&movq	(@T[1],@C[2]);
176	&psrlq	(@C[2],64-$rhotates[2][2]);
177	 &pxor	(@C[1],&QWP(@D[1],"esp"));
178	&psllq	(@T[1],$rhotates[2][2]);
179
180	 &movq	(@T[2],@C[1]);
181	 &psrlq	(@C[1],64-$rhotates[1][1]);
182	&por	(@C[2],@T[1]);		# C[2] = ROL64(A[2][2] ^ C[1], rhotates[2][2]);   /* D[2] */
183	 &psllq	(@T[2],$rhotates[1][1]);
184	&pxor	(@C[0],&QWP($A[0][0],"esi")); # /* rotate by 0 */  /* D[0] */
185	 &por	(@C[1],@T[2]);		# C[1] = ROL64(A[1][1] ^ D[1], rhotates[1][1]);
186
187sub Chi() {				######### regular Chi step
188    my ($y,$xrho) = @_;
189
190	&movq	(@T[0],@C[1]);
191	 &movq	(@T[1],@C[2]);
192	&pandn	(@T[0],@C[2]);
193	 &pandn	(@C[2],@C[3]);
194	&pxor	(@T[0],@C[0]);
195	 &pxor	(@C[2],@C[1]);
196	&pxor	(@T[0],&QWP(0,"ebx"))		if ($y == 0);
197	&lea	("ebx",&DWP(8,"ebx"))		if ($y == 0);
198
199	&movq	(@T[2],@C[3]);
200	&movq	(&QWP($A[$y][0],"edi"),@T[0]);	# R[0][0] = C[0] ^ (~C[1] & C[2]) ^ iotas[i];
201	 &movq	(@T[0],@C[4]);
202	&pandn	(@C[3],@C[4]);
203	 &pandn	(@C[4],@C[0]);
204	&pxor	(@C[3],@T[1]);
205	 &movq	(&QWP($A[$y][1],"edi"),@C[2]);	# R[0][1] = C[1] ^ (~C[2] & C[3]);
206	 &pxor	(@C[4],@T[2]);
207	  &movq	(@T[2],&QWP($A[0][$xrho],"esi"))	if (defined($xrho));
208
209	 &movq	(&QWP($A[$y][2],"edi"),@C[3]);	# R[0][2] = C[2] ^ (~C[3] & C[4]);
210	&pandn	(@C[0],@C[1]);
211	 &movq	(&QWP($A[$y][3],"edi"),@C[4]);	# R[0][3] = C[3] ^ (~C[4] & C[0]);
212	&pxor	(@C[0],@T[0]);
213	  &pxor	(@T[2],&QWP(@D[$xrho],"esp"))		if (defined($xrho));
214	&movq	(&QWP($A[$y][4],"edi"),@C[0]);	# R[0][4] = C[4] ^ (~C[0] & C[1]);
215}
216	&Chi	(0, 3);
217
218sub Rho() {				######### regular Rho step
219    my $x = shift;
220
221	#&movq	(@T[2],&QWP($A[0][$x],"esi"));	# moved to Chi
222	#&pxor	(@T[2],&QWP(@D[$x],"esp"));	# moved to Chi
223	&movq	(@C[0],@T[2]);
224	&psrlq	(@T[2],64-$rhotates[0][$x]);
225	 &movq	(@C[1],&QWP($A[1][($x+1)%5],"esi"));
226	&psllq	(@C[0],$rhotates[0][$x]);
227	 &pxor	(@C[1],&QWP(@D[($x+1)%5],"esp"));
228	&por	(@C[0],@T[2]);		# C[0] = ROL64(A[0][3] ^ D[3], rhotates[0][3]);
229
230	 &movq	(@T[1],@C[1]);
231	 &psrlq	(@C[1],64-$rhotates[1][($x+1)%5]);
232	&movq	(@C[2],&QWP($A[2][($x+2)%5],"esi"));
233	 &psllq	(@T[1],$rhotates[1][($x+1)%5]);
234	&pxor	(@C[2],&QWP(@D[($x+2)%5],"esp"));
235	 &por	(@C[1],@T[1]);		# C[1] = ROL64(A[1][4] ^ D[4], rhotates[1][4]);
236
237	&movq	(@T[2],@C[2]);
238	&psrlq	(@C[2],64-$rhotates[2][($x+2)%5]);
239	 &movq	(@C[3],&QWP($A[3][($x+3)%5],"esi"));
240	&psllq	(@T[2],$rhotates[2][($x+2)%5]);
241	 &pxor	(@C[3],&QWP(@D[($x+3)%5],"esp"));
242	&por	(@C[2],@T[2]);		# C[2] = ROL64(A[2][0] ^ D[0], rhotates[2][0]);
243
244	 &movq	(@T[0],@C[3]);
245	 &psrlq	(@C[3],64-$rhotates[3][($x+3)%5]);
246	&movq	(@C[4],&QWP($A[4][($x+4)%5],"esi"));
247	 &psllq	(@T[0],$rhotates[3][($x+3)%5]);
248	&pxor	(@C[4],&QWP(@D[($x+4)%5],"esp"));
249	 &por	(@C[3],@T[0]);		# C[3] = ROL64(A[3][1] ^ D[1], rhotates[3][1]);
250
251	&movq	(@T[1],@C[4]);
252	&psrlq	(@C[4],64-$rhotates[4][($x+4)%5]);
253	&psllq	(@T[1],$rhotates[4][($x+4)%5]);
254	&por	(@C[4],@T[1]);		# C[4] = ROL64(A[4][2] ^ D[2], rhotates[4][2]);
255}
256	&Rho	(3);	&Chi	(1, 1);
257	&Rho	(1);	&Chi	(2, 4);
258	&Rho	(4);	&Chi	(3, 2);
259	&Rho	(2);	###&Chi	(4);
260
261	&movq	(@T[0],@C[0]);		######### last Chi(4) is special
262	 &xor	("edi","esi");		# &xchg	("esi","edi");
263	&movq	(&QWP(@D[1],"esp"),@C[1]);
264	 &xor	("esi","edi");
265	 &xor	("edi","esi");
266
267	&movq	(@T[1],@C[1]);
268	 &movq	(@T[2],@C[2]);
269	&pandn	(@T[1],@C[2]);
270	 &pandn	(@T[2],@C[3]);
271	&pxor	(@C[0],@T[1]);
272	 &pxor	(@C[1],@T[2]);
273
274	&movq	(@T[1],@C[3]);
275	 &movq	(&QWP($A[4][0],"esi"),@C[0]);	# R[4][0] = C[0] ^= (~C[1] & C[2]);
276	&pandn	(@T[1],@C[4]);
277	 &movq	(&QWP($A[4][1],"esi"),@C[1]);	# R[4][1] = C[1] ^= (~C[2] & C[3]);
278	&pxor	(@C[2],@T[1]);
279	 &movq	(@T[2],@C[4]);
280	&movq	(&QWP($A[4][2],"esi"),@C[2]);	# R[4][2] = C[2] ^= (~C[3] & C[4]);
281
282	&pandn	(@T[2],@T[0]);
283	 &pandn	(@T[0],&QWP(@D[1],"esp"));
284	&pxor	(@C[3],@T[2]);
285	 &pxor	(@C[4],@T[0]);
286	&movq	(&QWP($A[4][3],"esi"),@C[3]);	# R[4][3] = C[3] ^= (~C[4] & D[0]);
287	&sub	("ecx",1);
288	 &movq	(&QWP($A[4][4],"esi"),@C[4]);	# R[4][4] = C[4] ^= (~D[0] & D[1]);
289	&jnz	(&label("loop"));
290
291	&lea	("ebx",&DWP(-192,"ebx"));	# rewind iotas
292	&ret	();
293&function_end_B("_KeccakF1600");
294
295&function_begin("KeccakF1600");
296	&mov	("esi",&wparam(0));
297	&mov	("ebp","esp");
298	&sub	("esp",240);
299	&call	(&label("pic_point"));
300    &set_label("pic_point");
301	&blindpop("ebx");
302	&lea	("ebx",&DWP(&label("iotas")."-".&label("pic_point"),"ebx"));
303	&and	("esp",-8);
304	&lea	("esi",&DWP(100,"esi"));	# size optimization
305	&lea	("edi",&DWP(8*5+100,"esp"));	# size optimization
306
307	&call	("_KeccakF1600");
308
309	&mov	("esp","ebp");
310	&emms	();
311&function_end("KeccakF1600");
312
313&function_begin("SHA3_absorb");
314	&mov	("esi",&wparam(0));		# A[][]
315	&mov	("eax",&wparam(1));		# inp
316	&mov	("ecx",&wparam(2));		# len
317	&mov	("edx",&wparam(3));		# bsz
318	&mov	("ebp","esp");
319	&sub	("esp",240+8);
320	&call	(&label("pic_point"));
321    &set_label("pic_point");
322	&blindpop("ebx");
323	&lea	("ebx",&DWP(&label("iotas")."-".&label("pic_point"),"ebx"));
324	&and	("esp",-8);
325
326	&mov	("edi","esi");
327	&lea	("esi",&DWP(100,"esi"));	# size optimization
328	&mov	(&DWP(-4,"ebp"),"edx");		# save bsz
329	&jmp	(&label("loop"));
330
331&set_label("loop",16);
332	&cmp	("ecx","edx");			# len < bsz?
333	&jc	(&label("absorbed"));
334
335	&shr	("edx",3);			# bsz /= 8
336&set_label("block");
337	&movq	("mm0",&QWP(0,"eax"));
338	&lea	("eax",&DWP(8,"eax"));
339	&pxor	("mm0",&QWP(0,"edi"));
340	&lea	("edi",&DWP(8,"edi"));
341	&sub	("ecx",8);			# len -= 8
342	&movq	(&QWP(-8,"edi"),"mm0");
343	&dec	("edx");			# bsz--
344	&jnz	(&label("block"));
345
346	&lea	("edi",&DWP(8*5+100,"esp"));	# size optimization
347	&mov	(&DWP(-8,"ebp"),"ecx");		# save len
348	&call	("_KeccakF1600");
349	&mov	("ecx",&DWP(-8,"ebp"));		# pull len
350	&mov	("edx",&DWP(-4,"ebp"));		# pull bsz
351	&lea	("edi",&DWP(-100,"esi"));
352	&jmp	(&label("loop"));
353
354&set_label("absorbed",16);
355	&mov	("eax","ecx");			# return value
356	&mov	("esp","ebp");
357	&emms	();
358&function_end("SHA3_absorb");
359
360&function_begin("SHA3_squeeze");
361	&mov	("esi",&wparam(0));		# A[][]
362	&mov	("eax",&wparam(1));		# out
363	&mov	("ecx",&wparam(2));		# len
364	&mov	("edx",&wparam(3));		# bsz
365	&mov	("ebp","esp");
366	&sub	("esp",240+8);
367	&call	(&label("pic_point"));
368    &set_label("pic_point");
369	&blindpop("ebx");
370	&lea	("ebx",&DWP(&label("iotas")."-".&label("pic_point"),"ebx"));
371	&and	("esp",-8);
372
373	&shr	("edx",3);			# bsz /= 8
374	&mov	("edi","esi");
375	&lea	("esi",&DWP(100,"esi"));	# size optimization
376	&mov	(&DWP(-4,"ebp"),"edx");		# save bsz
377	&jmp	(&label("loop"));
378
379&set_label("loop",16);
380	&cmp	("ecx",8);			# len < 8?
381	&jc	(&label("tail"));
382
383	&movq	("mm0",&QWP(0,"edi"));
384	&lea	("edi",&DWP(8,"edi"));
385	&movq	(&QWP(0,"eax"),"mm0");
386	&lea	("eax",&DWP(8,"eax"));
387	&sub	("ecx",8);			# len -= 8
388	&jz	(&label("done"));
389
390	&dec	("edx");			# bsz--
391	&jnz	(&label("loop"));
392
393	&lea	("edi",&DWP(8*5+100,"esp"));	# size optimization
394	&mov	(&DWP(-8,"ebp"),"ecx");		# save len
395	&call	("_KeccakF1600");
396	&mov	("ecx",&DWP(-8,"ebp"));		# pull len
397	&mov	("edx",&DWP(-4,"ebp"));		# pull bsz
398	&lea	("edi",&DWP(-100,"esi"));
399	&jmp	(&label("loop"));
400
401&set_label("tail",16);
402	&mov	("esi","edi");
403	&mov	("edi","eax");
404	&data_word("0xA4F39066");		# rep movsb
405
406&set_label("done");
407	&mov	("esp","ebp");
408	&emms	();
409&function_end("SHA3_squeeze");
410
411&set_label("iotas",32);
412	&data_word(0x00000001,0x00000000);
413	&data_word(0x00008082,0x00000000);
414	&data_word(0x0000808a,0x80000000);
415	&data_word(0x80008000,0x80000000);
416	&data_word(0x0000808b,0x00000000);
417	&data_word(0x80000001,0x00000000);
418	&data_word(0x80008081,0x80000000);
419	&data_word(0x00008009,0x80000000);
420	&data_word(0x0000008a,0x00000000);
421	&data_word(0x00000088,0x00000000);
422	&data_word(0x80008009,0x00000000);
423	&data_word(0x8000000a,0x00000000);
424	&data_word(0x8000808b,0x00000000);
425	&data_word(0x0000008b,0x80000000);
426	&data_word(0x00008089,0x80000000);
427	&data_word(0x00008003,0x80000000);
428	&data_word(0x00008002,0x80000000);
429	&data_word(0x00000080,0x80000000);
430	&data_word(0x0000800a,0x00000000);
431	&data_word(0x8000000a,0x80000000);
432	&data_word(0x80008081,0x80000000);
433	&data_word(0x00008080,0x80000000);
434	&data_word(0x80000001,0x00000000);
435	&data_word(0x80008008,0x80000000);
436&asciz("Keccak-1600 absorb and squeeze for MMX, CRYPTOGAMS by <appro\@openssl.org>");
437
438&asm_finish();
439
440close STDOUT;
441