1#!/usr/bin/perl -w
2#
3# MD5 optimized for AMD64.
4#
5# Author: Marc Bevand <bevand_m (at) epita.fr>
6# Licence: I hereby disclaim the copyright on this code and place it
7# in the public domain.
8#
9
10#
11# The following is Marc Bevand's MD5 implementation optimized for
12# AMD64.  It has been lifted intact, except for changing the comment
13# character and adding comments.
14#
15# typedef struct {
16#	uint32_t state[4];	/* state (ABCD) */
17#	uint32_t count[2];	/* number of bits, modulo 2^64 (lsb first) */
18#	union	{
19#		uint8_t		buf8[64];	/* undigested input */
20#		uint32_t	buf32[16];	/* realigned input */
21#		} buf_un;
22#	} MD5_CTX;
23#
24# void md5_block_asm_host_order(MD5_CTX *ctx, const void *inpp,
25#        unsigned int input_length_in_blocks);
26#
27# Registers used:
28#	rax  A		r8  old A
29#	rbx  B		r9  old B
30#	rcx  C		r10 tmp
31#	rdx  D		r11 tmp
32#	rsi  ptr	r12 tmp
33#	rdi  end	r13 -
34#	rbp  -		r14 old C
35#	rsp  stack	r15 old D
36#
37
38use strict;
39my $code;
40
41
42# round1_step() does:
43#   dst = x + ((dst + F(x,y,z) + X[k] + T_i) <<< s)
44#   %r10d = X[k_next]
45#   %r11d = z' (copy of z for the next step)
46# Each round1_step() takes about 5.3 clocks (9 instructions, 1.7 IPC)
47sub round1_step
48{
49    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
50    $code .= "	mov	0*4(%rsi),	%r10d		/* (NEXT STEP) X[0] */\n" if ($pos == -1);
51    $code .= "	mov	%edx,		%r11d		/* (NEXT STEP) z' = %edx */\n" if ($pos == -1);
52
53    $code .= <<EOF;
54	xor	$y,		%r11d		/* y ^ ... */
55	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
56	and	$x,		%r11d		/* x & ... */
57	xor	$z,		%r11d		/* z ^ ... */
58	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
59	add	%r11d,		$dst		/* dst += ... */
60	rol	\$$s,		$dst		/* dst <<< s */
61	mov	$y,		%r11d		/* (NEXT STEP) z' = $y */
62	add	$x,		$dst		/* dst += x */
63EOF
64}
65
66# round2_step() does:
67#   dst = x + ((dst + G(x,y,z) + X[k] + T_i) <<< s)
68#   %r10d = X[k_next]
69#   %r11d = z' (copy of z for the next step)
70#   %r12d = z' (copy of z for the next step)
71# Each round2_step() takes about 5.4 clocks (11 instructions, 2.0 IPC)
72sub round2_step
73{
74    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
75    $code .= "	mov	1*4(%rsi),	%r10d		/* (NEXT STEP) X[1] */\n" if ($pos == -1);
76    $code .= "	mov	%edx,		%r11d		/* (NEXT STEP) z' = %edx */\n" if ($pos == -1);
77    $code .= "	mov	%edx,		%r12d		/* (NEXT STEP) z' = %edx */\n" if ($pos == -1);
78
79    $code .= <<EOF;
80	not	%r11d				/* not z */
81	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
82	and	$x,		%r12d		/* x & z */
83	and	$y,		%r11d		/* y & (not z) */
84	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
85	or	%r11d,		%r12d		/* (y & (not z)) | (x & z) */
86	mov	$y,		%r11d		/* (NEXT STEP) z' = $y */
87	add	%r12d,		$dst		/* dst += ... */
88	mov	$y,		%r12d		/* (NEXT STEP) z' = $y */
89	rol	\$$s,		$dst		/* dst <<< s */
90	add	$x,		$dst		/* dst += x */
91EOF
92}
93
94# round3_step() does:
95#   dst = x + ((dst + H(x,y,z) + X[k] + T_i) <<< s)
96#   %r10d = X[k_next]
97#   %r11d = y' (copy of y for the next step)
98# Each round3_step() takes about 4.2 clocks (8 instructions, 1.9 IPC)
99sub round3_step
100{
101    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
102    $code .= "	mov	5*4(%rsi),	%r10d		/* (NEXT STEP) X[5] */\n" if ($pos == -1);
103    $code .= "	mov	%ecx,		%r11d		/* (NEXT STEP) y' = %ecx */\n" if ($pos == -1);
104
105    $code .= <<EOF;
106	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
107	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
108	xor	$z,		%r11d		/* z ^ ... */
109	xor	$x,		%r11d		/* x ^ ... */
110	add	%r11d,		$dst		/* dst += ... */
111	rol	\$$s,		$dst		/* dst <<< s */
112	mov	$x,		%r11d		/* (NEXT STEP) y' = $x */
113	add	$x,		$dst		/* dst += x */
114EOF
115}
116
117# round4_step() does:
118#   dst = x + ((dst + I(x,y,z) + X[k] + T_i) <<< s)
119#   %r10d = X[k_next]
120#   %r11d = not z' (copy of not z for the next step)
121# Each round4_step() takes about 5.2 clocks (9 instructions, 1.7 IPC)
122sub round4_step
123{
124    my ($pos, $dst, $x, $y, $z, $k_next, $T_i, $s) = @_;
125    $code .= "	mov	0*4(%rsi),	%r10d		/* (NEXT STEP) X[0] */\n" if ($pos == -1);
126    $code .= "	mov	\$0xffffffff,	%r11d\n" if ($pos == -1);
127    $code .= "	xor	%edx,		%r11d		/* (NEXT STEP) not z' = not %edx*/\n"
128    if ($pos == -1);
129
130    $code .= <<EOF;
131	lea	$T_i($dst,%r10d),$dst		/* Const + dst + ... */
132	or	$x,		%r11d		/* x | ... */
133	xor	$y,		%r11d		/* y ^ ... */
134	add	%r11d,		$dst		/* dst += ... */
135	mov	$k_next*4(%rsi),%r10d		/* (NEXT STEP) X[$k_next] */
136	mov	\$0xffffffff,	%r11d
137	rol	\$$s,		$dst		/* dst <<< s */
138	xor	$y,		%r11d		/* (NEXT STEP) not z' = not $y */
139	add	$x,		$dst		/* dst += x */
140EOF
141}
142
143
144#
145# Execution begins here.
146#
147
148my $output = shift;
149open STDOUT,">$output" or die "can't open $output: $!";
150
151$code .= <<EOF;
152#if defined(lint) || defined(__lint)
153#include <sys/md5.h>
154
155/* ARGSUSED */
156void md5_block_asm_host_order(MD5_CTX *ctx, const void *inpp,
157    unsigned int input_length_in_blocks)
158{
159}
160
161#else
162#include <sys/asm_linkage.h>
163
164	ENTRY_NP(md5_block_asm_host_order)
165	push	%rbp
166	push	%rbx
167	push	%r12
168	push	%r13
169	push	%r14
170	push	%r15
171
172	/ rdi = arg #1 (ctx, MD5_CTX pointer)
173	/ rsi = arg #2 (ptr, data pointer)
174	/ rdx = arg #3 (nbr, number of 64-byte blocks to process)
175	mov	%rdi,		%rbp	/ rbp = ctx
176	shl	\$6,		%rdx	/ rdx = nbr in bytes
177	lea	(%rsi,%rdx),	%rdi	/ rdi = end
178	mov	0*4(%rbp),	%eax	/ eax = ctx->A
179	mov	1*4(%rbp),	%ebx	/ ebx = ctx->B
180	mov	2*4(%rbp),	%ecx	/ ecx = ctx->C
181	mov	3*4(%rbp),	%edx	/ edx = ctx->D
182	push	%rbp			/ save ctx
183	/ end is 'rdi'
184	/ ptr is 'rsi'
185	/ A is 'eax'
186	/ B is 'ebx'
187	/ C is 'ecx'
188	/ D is 'edx'
189
190	cmp	%rdi,		%rsi		/ cmp end with ptr
191	je	1f				/ jmp if ptr == end
192
193	/ BEGIN of loop over 64-byte blocks
1942:	/ save old values of A, B, C, D
195	mov	%eax,		%r8d
196	mov	%ebx,		%r9d
197	mov	%ecx,		%r14d
198	mov	%edx,		%r15d
199EOF
200round1_step(-1,'%eax','%ebx','%ecx','%edx', '1','0xd76aa478', '7');
201round1_step( 0,'%edx','%eax','%ebx','%ecx', '2','0xe8c7b756','12');
202round1_step( 0,'%ecx','%edx','%eax','%ebx', '3','0x242070db','17');
203round1_step( 0,'%ebx','%ecx','%edx','%eax', '4','0xc1bdceee','22');
204round1_step( 0,'%eax','%ebx','%ecx','%edx', '5','0xf57c0faf', '7');
205round1_step( 0,'%edx','%eax','%ebx','%ecx', '6','0x4787c62a','12');
206round1_step( 0,'%ecx','%edx','%eax','%ebx', '7','0xa8304613','17');
207round1_step( 0,'%ebx','%ecx','%edx','%eax', '8','0xfd469501','22');
208round1_step( 0,'%eax','%ebx','%ecx','%edx', '9','0x698098d8', '7');
209round1_step( 0,'%edx','%eax','%ebx','%ecx','10','0x8b44f7af','12');
210round1_step( 0,'%ecx','%edx','%eax','%ebx','11','0xffff5bb1','17');
211round1_step( 0,'%ebx','%ecx','%edx','%eax','12','0x895cd7be','22');
212round1_step( 0,'%eax','%ebx','%ecx','%edx','13','0x6b901122', '7');
213round1_step( 0,'%edx','%eax','%ebx','%ecx','14','0xfd987193','12');
214round1_step( 0,'%ecx','%edx','%eax','%ebx','15','0xa679438e','17');
215round1_step( 1,'%ebx','%ecx','%edx','%eax', '0','0x49b40821','22');
216
217round2_step(-1,'%eax','%ebx','%ecx','%edx', '6','0xf61e2562', '5');
218round2_step( 0,'%edx','%eax','%ebx','%ecx','11','0xc040b340', '9');
219round2_step( 0,'%ecx','%edx','%eax','%ebx', '0','0x265e5a51','14');
220round2_step( 0,'%ebx','%ecx','%edx','%eax', '5','0xe9b6c7aa','20');
221round2_step( 0,'%eax','%ebx','%ecx','%edx','10','0xd62f105d', '5');
222round2_step( 0,'%edx','%eax','%ebx','%ecx','15', '0x2441453', '9');
223round2_step( 0,'%ecx','%edx','%eax','%ebx', '4','0xd8a1e681','14');
224round2_step( 0,'%ebx','%ecx','%edx','%eax', '9','0xe7d3fbc8','20');
225round2_step( 0,'%eax','%ebx','%ecx','%edx','14','0x21e1cde6', '5');
226round2_step( 0,'%edx','%eax','%ebx','%ecx', '3','0xc33707d6', '9');
227round2_step( 0,'%ecx','%edx','%eax','%ebx', '8','0xf4d50d87','14');
228round2_step( 0,'%ebx','%ecx','%edx','%eax','13','0x455a14ed','20');
229round2_step( 0,'%eax','%ebx','%ecx','%edx', '2','0xa9e3e905', '5');
230round2_step( 0,'%edx','%eax','%ebx','%ecx', '7','0xfcefa3f8', '9');
231round2_step( 0,'%ecx','%edx','%eax','%ebx','12','0x676f02d9','14');
232round2_step( 1,'%ebx','%ecx','%edx','%eax', '0','0x8d2a4c8a','20');
233
234round3_step(-1,'%eax','%ebx','%ecx','%edx', '8','0xfffa3942', '4');
235round3_step( 0,'%edx','%eax','%ebx','%ecx','11','0x8771f681','11');
236round3_step( 0,'%ecx','%edx','%eax','%ebx','14','0x6d9d6122','16');
237round3_step( 0,'%ebx','%ecx','%edx','%eax', '1','0xfde5380c','23');
238round3_step( 0,'%eax','%ebx','%ecx','%edx', '4','0xa4beea44', '4');
239round3_step( 0,'%edx','%eax','%ebx','%ecx', '7','0x4bdecfa9','11');
240round3_step( 0,'%ecx','%edx','%eax','%ebx','10','0xf6bb4b60','16');
241round3_step( 0,'%ebx','%ecx','%edx','%eax','13','0xbebfbc70','23');
242round3_step( 0,'%eax','%ebx','%ecx','%edx', '0','0x289b7ec6', '4');
243round3_step( 0,'%edx','%eax','%ebx','%ecx', '3','0xeaa127fa','11');
244round3_step( 0,'%ecx','%edx','%eax','%ebx', '6','0xd4ef3085','16');
245round3_step( 0,'%ebx','%ecx','%edx','%eax', '9', '0x4881d05','23');
246round3_step( 0,'%eax','%ebx','%ecx','%edx','12','0xd9d4d039', '4');
247round3_step( 0,'%edx','%eax','%ebx','%ecx','15','0xe6db99e5','11');
248round3_step( 0,'%ecx','%edx','%eax','%ebx', '2','0x1fa27cf8','16');
249round3_step( 1,'%ebx','%ecx','%edx','%eax', '0','0xc4ac5665','23');
250
251round4_step(-1,'%eax','%ebx','%ecx','%edx', '7','0xf4292244', '6');
252round4_step( 0,'%edx','%eax','%ebx','%ecx','14','0x432aff97','10');
253round4_step( 0,'%ecx','%edx','%eax','%ebx', '5','0xab9423a7','15');
254round4_step( 0,'%ebx','%ecx','%edx','%eax','12','0xfc93a039','21');
255round4_step( 0,'%eax','%ebx','%ecx','%edx', '3','0x655b59c3', '6');
256round4_step( 0,'%edx','%eax','%ebx','%ecx','10','0x8f0ccc92','10');
257round4_step( 0,'%ecx','%edx','%eax','%ebx', '1','0xffeff47d','15');
258round4_step( 0,'%ebx','%ecx','%edx','%eax', '8','0x85845dd1','21');
259round4_step( 0,'%eax','%ebx','%ecx','%edx','15','0x6fa87e4f', '6');
260round4_step( 0,'%edx','%eax','%ebx','%ecx', '6','0xfe2ce6e0','10');
261round4_step( 0,'%ecx','%edx','%eax','%ebx','13','0xa3014314','15');
262round4_step( 0,'%ebx','%ecx','%edx','%eax', '4','0x4e0811a1','21');
263round4_step( 0,'%eax','%ebx','%ecx','%edx','11','0xf7537e82', '6');
264round4_step( 0,'%edx','%eax','%ebx','%ecx', '2','0xbd3af235','10');
265round4_step( 0,'%ecx','%edx','%eax','%ebx', '9','0x2ad7d2bb','15');
266round4_step( 1,'%ebx','%ecx','%edx','%eax', '0','0xeb86d391','21');
267$code .= <<EOF;
268	/ add old values of A, B, C, D
269	add	%r8d,	%eax
270	add	%r9d,	%ebx
271	add	%r14d,	%ecx
272	add	%r15d,	%edx
273
274	/ loop control
275	add	\$64,		%rsi		/ ptr += 64
276	cmp	%rdi,		%rsi		/ cmp end with ptr
277	jb	2b				/ jmp if ptr < end
278	/ END of loop over 64-byte blocks
279
2801:	pop	%rbp				/ restore ctx
281	mov	%eax,		0*4(%rbp)	/ ctx->A = A
282	mov	%ebx,		1*4(%rbp)	/ ctx->B = B
283	mov	%ecx,		2*4(%rbp)	/ ctx->C = C
284	mov	%edx,		3*4(%rbp)	/ ctx->D = D
285
286	pop	%r15
287	pop	%r14
288	pop	%r13
289	pop	%r12
290	pop	%rbx
291	pop	%rbp
292	ret
293	SET_SIZE(md5_block_asm_host_order)
294
295#endif /* lint || __lint */
296EOF
297
298print $code;
299