sha512-x86_64.pl revision e45f106cb6b47af1f21efe76e933bdea2f5dd1ca 
1#!/usr/bin/env perl
2#
3# ====================================================================
4# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5# project. Rights for redistribution and usage in source and binary
6# forms are granted according to the OpenSSL license.
7# ====================================================================
8#
9# sha256/512_block procedure for x86_64.
10#
11# 40% improvement over compiler-generated code on Opteron. On EM64T
12# sha256 was observed to run >80% faster and sha512 - >40%. No magical
13# tricks, just straight implementation... I really wonder why gcc
14# [being armed with inline assembler] fails to generate as fast code.
15# The only thing which is cool about this module is that it's very
16# same instruction sequence used for both SHA-256 and SHA-512. In
17# former case the instructions operate on 32-bit operands, while in
18# latter - on 64-bit ones. All I had to do is to get one flavor right,
19# the other one passed the test right away:-)
20#
21# sha256_block runs in ~1005 cycles on Opteron, which gives you
22# asymptotic performance of 64*1000/1005=63.7MBps times CPU clock
23# frequency in GHz. sha512_block runs in ~1275 cycles, which results
24# in 128*1000/1275=100MBps per GHz. Is there room for improvement?
25# Well, if you compare it to IA-64 implementation, which maintains
26# X[16] in register bank[!], tends to 4 instructions per CPU clock
27# cycle and runs in 1003 cycles, 1275 is very good result for 3-way
28# issue Opteron pipeline and X[16] maintained in memory. So that *if*
29# there is a way to improve it, *then* the only way would be to try to
30# offload X[16] updates to SSE unit, but that would require "deeper"
31# loop unroll, which in turn would naturally cause size blow-up, not
32# to mention increased complexity! And once again, only *if* it's
33# actually possible to noticeably improve overall ILP, instruction
34# level parallelism, on a given CPU implementation in this case.
35#
36# Special note on Intel EM64T. While Opteron CPU exhibits perfect
37# perfromance ratio of 1.5 between 64- and 32-bit flavors [see above],
38# [currently available] EM64T CPUs apparently are far from it. On the
39# contrary, 64-bit version, sha512_block, is ~30% *slower* than 32-bit
40# sha256_block:-( This is presumably because 64-bit shifts/rotates
41# apparently are not atomic instructions, but implemented in microcode.
42
43$output=shift;
44
45$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
46( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
47( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
48die "can't locate x86_64-xlate.pl";
49
50open STDOUT,"| $^X $xlate $output";
51
52if ($output =~ /512/) {
53	$func="sha512_block_data_order";
54	$TABLE="K512";
55	$SZ=8;
56	@ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%rax","%rbx","%rcx","%rdx",
57					"%r8", "%r9", "%r10","%r11");
58	($T1,$a0,$a1,$a2)=("%r12","%r13","%r14","%r15");
59	@Sigma0=(28,34,39);
60	@Sigma1=(14,18,41);
61	@sigma0=(1,  8, 7);
62	@sigma1=(19,61, 6);
63	$rounds=80;
64} else {
65	$func="sha256_block_data_order";
66	$TABLE="K256";
67	$SZ=4;
68	@ROT=($A,$B,$C,$D,$E,$F,$G,$H)=("%eax","%ebx","%ecx","%edx",
69					"%r8d","%r9d","%r10d","%r11d");
70	($T1,$a0,$a1,$a2)=("%r12d","%r13d","%r14d","%r15d");
71	@Sigma0=( 2,13,22);
72	@Sigma1=( 6,11,25);
73	@sigma0=( 7,18, 3);
74	@sigma1=(17,19,10);
75	$rounds=64;
76}
77
78$ctx="%rdi";	# 1st arg
79$round="%rdi";	# zaps $ctx
80$inp="%rsi";	# 2nd arg
81$Tbl="%rbp";
82
83$_ctx="16*$SZ+0*8(%rsp)";
84$_inp="16*$SZ+1*8(%rsp)";
85$_end="16*$SZ+2*8(%rsp)";
86$_rsp="16*$SZ+3*8(%rsp)";
87$framesz="16*$SZ+4*8";
88
89
90sub ROUND_00_15()
91{ my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
92
93$code.=<<___;
94	mov	$e,$a0
95	mov	$e,$a1
96	mov	$f,$a2
97
98	ror	\$$Sigma1[0],$a0
99	ror	\$$Sigma1[1],$a1
100	xor	$g,$a2			# f^g
101
102	xor	$a1,$a0
103	ror	\$`$Sigma1[2]-$Sigma1[1]`,$a1
104	and	$e,$a2			# (f^g)&e
105	mov	$T1,`$SZ*($i&0xf)`(%rsp)
106
107	xor	$a1,$a0			# Sigma1(e)
108	xor	$g,$a2			# Ch(e,f,g)=((f^g)&e)^g
109	add	$h,$T1			# T1+=h
110
111	mov	$a,$h
112	add	$a0,$T1			# T1+=Sigma1(e)
113
114	add	$a2,$T1			# T1+=Ch(e,f,g)
115	mov	$a,$a0
116	mov	$a,$a1
117
118	ror	\$$Sigma0[0],$h
119	ror	\$$Sigma0[1],$a0
120	mov	$a,$a2
121	add	($Tbl,$round,$SZ),$T1	# T1+=K[round]
122
123	xor	$a0,$h
124	ror	\$`$Sigma0[2]-$Sigma0[1]`,$a0
125	or	$c,$a1			# a|c
126
127	xor	$a0,$h			# h=Sigma0(a)
128	and	$c,$a2			# a&c
129	add	$T1,$d			# d+=T1
130
131	and	$b,$a1			# (a|c)&b
132	add	$T1,$h			# h+=T1
133
134	or	$a2,$a1			# Maj(a,b,c)=((a|c)&b)|(a&c)
135	lea	1($round),$round	# round++
136
137	add	$a1,$h			# h+=Maj(a,b,c)
138___
139}
140
141sub ROUND_16_XX()
142{ my ($i,$a,$b,$c,$d,$e,$f,$g,$h) = @_;
143
144$code.=<<___;
145	mov	`$SZ*(($i+1)&0xf)`(%rsp),$a0
146	mov	`$SZ*(($i+14)&0xf)`(%rsp),$T1
147
148	mov	$a0,$a2
149
150	shr	\$$sigma0[2],$a0
151	ror	\$$sigma0[0],$a2
152
153	xor	$a2,$a0
154	ror	\$`$sigma0[1]-$sigma0[0]`,$a2
155
156	xor	$a2,$a0			# sigma0(X[(i+1)&0xf])
157	mov	$T1,$a1
158
159	shr	\$$sigma1[2],$T1
160	ror	\$$sigma1[0],$a1
161
162	xor	$a1,$T1
163	ror	\$`$sigma1[1]-$sigma1[0]`,$a1
164
165	xor	$a1,$T1			# sigma1(X[(i+14)&0xf])
166
167	add	$a0,$T1
168
169	add	`$SZ*(($i+9)&0xf)`(%rsp),$T1
170
171	add	`$SZ*($i&0xf)`(%rsp),$T1
172___
173	&ROUND_00_15(@_);
174}
175
176$code=<<___;
177.text
178
179.globl	$func
180.type	$func,\@function,4
181.align	16
182$func:
183	push	%rbx
184	push	%rbp
185	push	%r12
186	push	%r13
187	push	%r14
188	push	%r15
189	mov	%rsp,%rbp		# copy %rsp
190	shl	\$4,%rdx		# num*16
191	sub	\$$framesz,%rsp
192	lea	($inp,%rdx,$SZ),%rdx	# inp+num*16*$SZ
193	and	\$-64,%rsp		# align stack frame
194	mov	$ctx,$_ctx		# save ctx, 1st arg
195	mov	$inp,$_inp		# save inp, 2nd arh
196	mov	%rdx,$_end		# save end pointer, "3rd" arg
197	mov	%rbp,$_rsp		# save copy of %rsp
198
199	.picmeup $Tbl
200	lea	$TABLE-.($Tbl),$Tbl
201
202	mov	$SZ*0($ctx),$A
203	mov	$SZ*1($ctx),$B
204	mov	$SZ*2($ctx),$C
205	mov	$SZ*3($ctx),$D
206	mov	$SZ*4($ctx),$E
207	mov	$SZ*5($ctx),$F
208	mov	$SZ*6($ctx),$G
209	mov	$SZ*7($ctx),$H
210	jmp	.Lloop
211
212.align	16
213.Lloop:
214	xor	$round,$round
215___
216	for($i=0;$i<16;$i++) {
217		$code.="	mov	$SZ*$i($inp),$T1\n";
218		$code.="	bswap	$T1\n";
219		&ROUND_00_15($i,@ROT);
220		unshift(@ROT,pop(@ROT));
221	}
222$code.=<<___;
223	jmp	.Lrounds_16_xx
224.align	16
225.Lrounds_16_xx:
226___
227	for(;$i<32;$i++) {
228		&ROUND_16_XX($i,@ROT);
229		unshift(@ROT,pop(@ROT));
230	}
231
232$code.=<<___;
233	cmp	\$$rounds,$round
234	jb	.Lrounds_16_xx
235
236	mov	$_ctx,$ctx
237	lea	16*$SZ($inp),$inp
238
239	add	$SZ*0($ctx),$A
240	add	$SZ*1($ctx),$B
241	add	$SZ*2($ctx),$C
242	add	$SZ*3($ctx),$D
243	add	$SZ*4($ctx),$E
244	add	$SZ*5($ctx),$F
245	add	$SZ*6($ctx),$G
246	add	$SZ*7($ctx),$H
247
248	cmp	$_end,$inp
249
250	mov	$A,$SZ*0($ctx)
251	mov	$B,$SZ*1($ctx)
252	mov	$C,$SZ*2($ctx)
253	mov	$D,$SZ*3($ctx)
254	mov	$E,$SZ*4($ctx)
255	mov	$F,$SZ*5($ctx)
256	mov	$G,$SZ*6($ctx)
257	mov	$H,$SZ*7($ctx)
258	jb	.Lloop
259
260	mov	$_rsp,%rsp
261	pop	%r15
262	pop	%r14
263	pop	%r13
264	pop	%r12
265	pop	%rbp
266	pop	%rbx
267
268	ret
269.size	$func,.-$func
270___
271
272if ($SZ==4) {
273$code.=<<___;
274.align	64
275.type	$TABLE,\@object
276$TABLE:
277	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
278	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
279	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
280	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
281	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
282	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
283	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
284	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
285	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
286	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
287	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
288	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
289	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
290	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
291	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
292	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
293___
294} else {
295$code.=<<___;
296.align	64
297.type	$TABLE,\@object
298$TABLE:
299	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
300	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
301	.quad	0x3956c25bf348b538,0x59f111f1b605d019
302	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
303	.quad	0xd807aa98a3030242,0x12835b0145706fbe
304	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
305	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
306	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
307	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
308	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
309	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
310	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
311	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
312	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
313	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
314	.quad	0x06ca6351e003826f,0x142929670a0e6e70
315	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
316	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
317	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
318	.quad	0x81c2c92e47edaee6,0x92722c851482353b
319	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
320	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
321	.quad	0xd192e819d6ef5218,0xd69906245565a910
322	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
323	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
324	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
325	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
326	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
327	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
328	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
329	.quad	0x90befffa23631e28,0xa4506cebde82bde9
330	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
331	.quad	0xca273eceea26619c,0xd186b8c721c0c207
332	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
333	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
334	.quad	0x113f9804bef90dae,0x1b710b35131c471b
335	.quad	0x28db77f523047d84,0x32caab7b40c72493
336	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
337	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
338	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
339___
340}
341
342$code =~ s/\`([^\`]*)\`/eval $1/gem;
343print $code;
344close STDOUT;
345