[PATCH] ARM optimized SHA1 implementation
This is my ARM assembly SHA1 implementation for GIT. It is approximately 50% faster than the generic C version. On an XScale processor running at 400MHz: generic C version: 9.8 MB/s my version: 14.5 MB/s It's not that I expect a lot of big GIT users on ARM, but I stillknow about one important ARM user that might benefit from it, and writing that code was fun. I also reworked the makefile a bit so any optimized SHA1 implementations is used regardless of whether NO_OPENSSL is defined or not. Signed-off-by: Nicolas Pitre <nico@cam.org> Signed-off-by: Junio C Hamano <junkio@cox.net>
Nicolas Pitre committed
Sep 20, 2005 at 12:27 UTC
7c6ef2f2142f0a4c22b5505d95553cac17867a21
4 files changed
+308
-12
Makefile
+24
-12
@@ -14,6 +14,9 @@
14
# Define PPC_SHA1 environment variable when running make to make use of
15
# a bundled SHA1 routine optimized for PowerPC.
16
#
17
+# Define ARM_SHA1 environment variable when running make to make use of
18
+# a bundled SHA1 routine optimized for ARM.
19
+#
20
# Define NEEDS_SSL_WITH_CRYPTO if you need -lcrypto with -lssl (Darwin).
21
#
22
# Define NEEDS_LIBICONV if linking with libc is not enough (Darwin).
@@ -162,6 +165,9 @@ ifeq ($(shell uname -s),SunOS)
165
NEEDS_NSL = YesPlease
166
PLATFORM_DEFINES += -D__EXTENSIONS__
167
endif
168
+ifneq (,$(findstring arm,$(shell uname -m)))
169
+ ARM_SHA1 = YesPlease
170
+endif
171
172
ifndef SHELL_PATH
173
SHELL_PATH = /bin/sh
@@ -191,18 +197,6 @@ ifdef NEEDS_LIBICONV
197
else
198
LIB_4_ICONV =
199
endif
194
-ifdef MOZILLA_SHA1
195
- SHA1_HEADER = "mozilla-sha1/sha1.h"
196
- LIB_OBJS += mozilla-sha1/sha1.o
197
-else
198
- ifdef PPC_SHA1
199
- SHA1_HEADER = "ppc/sha1.h"
200
- LIB_OBJS += ppc/sha1.o ppc/sha1ppc.o
201
- else
202
- SHA1_HEADER = <openssl/sha.h>
203
- LIBS += $(LIB_4_CRYPTO)
204
- endif
205
-endif
200
ifdef NEEDS_SOCKET
201
LIBS += -lsocket
202
SIMPLE_LIB += -lsocket
@@ -216,6 +210,24 @@ ifdef NO_STRCASESTR
210
LIB_OBJS += compat/strcasestr.o
211
endif
212
213
+ifdef PPC_SHA1
214
+ SHA1_HEADER = "ppc/sha1.h"
215
+ LIB_OBJS += ppc/sha1.o ppc/sha1ppc.o
216
+else
217
+ifdef ARM_SHA1
218
+ SHA1_HEADER = "arm/sha1.h"
219
+ LIB_OBJS += arm/sha1.o arm/sha1_arm.o
220
+else
221
+ifdef MOZILLA_SHA1
222
+ SHA1_HEADER = "mozilla-sha1/sha1.h"
223
+ LIB_OBJS += mozilla-sha1/sha1.o
224
+else
225
+ SHA1_HEADER = <openssl/sha.h>
226
+ LIBS += $(LIB_4_CRYPTO)
227
+endif
228
+endif
229
+endif
230
+
231
DEFINES += '-DSHA1_HEADER=$(SHA1_HEADER)'
232
233
SCRIPTS = $(patsubst %.sh,%,$(SCRIPT_SH)) \
arm/sha1.c
new
+82
@@ -0,0 +1,82 @@
1
+/*
2
+ * SHA-1 implementation optimized for ARM
3
+ *
4
+ * Copyright: (C) 2005 by Nicolas Pitre <nico@cam.org>
5
+ * Created: September 17, 2005
6
+ */
7
+
8
+#include <string.h>
9
+#include "sha1.h"
10
+
11
+extern void sha_transform(uint32_t *hash, const unsigned char *data, uint32_t *W);
12
+
13
+void SHA1_Init(SHA_CTX *c)
14
+{
15
+ c->len = 0;
16
+ c->hash[0] = 0x67452301;
17
+ c->hash[1] = 0xefcdab89;
18
+ c->hash[2] = 0x98badcfe;
19
+ c->hash[3] = 0x10325476;
20
+ c->hash[4] = 0xc3d2e1f0;
21
+}
22
+
23
+void SHA1_Update(SHA_CTX *c, const void *p, unsigned long n)
24
+{
25
+ uint32_t workspace[80];
26
+ unsigned int partial;
27
+ unsigned long done;
28
+
29
+ partial = c->len & 0x3f;
30
+ c->len += n;
31
+ if ((partial + n) >= 64) {
32
+ if (partial) {
33
+ done = 64 - partial;
34
+ memcpy(c->buffer + partial, p, done);
35
+ sha_transform(c->hash, c->buffer, workspace);
36
+ partial = 0;
37
+ } else
38
+ done = 0;
39
+ while (n >= done + 64) {
40
+ sha_transform(c->hash, p + done, workspace);
41
+ done += 64;
42
+ }
43
+ } else
44
+ done = 0;
45
+ if (n - done)
46
+ memcpy(c->buffer + partial, p + done, n - done);
47
+}
48
+
49
+void SHA1_Final(unsigned char *hash, SHA_CTX *c)
50
+{
51
+ uint64_t bitlen;
52
+ uint32_t bitlen_hi, bitlen_lo;
53
+ unsigned int i, offset, padlen;
54
+ unsigned char bits[8];
55
+ static const unsigned char padding[64] = { 0x80, };
56
+
57
+ bitlen = c->len << 3;
58
+ offset = c->len & 0x3f;
59
+ padlen = ((offset < 56) ? 56 : (64 + 56)) - offset;
60
+ SHA1_Update(c, padding, padlen);
61
+
62
+ bitlen_hi = bitlen >> 32;
63
+ bitlen_lo = bitlen & 0xffffffff;
64
+ bits[0] = bitlen_hi >> 24;
65
+ bits[1] = bitlen_hi >> 16;
66
+ bits[2] = bitlen_hi >> 8;
67
+ bits[3] = bitlen_hi;
68
+ bits[4] = bitlen_lo >> 24;
69
+ bits[5] = bitlen_lo >> 16;
70
+ bits[6] = bitlen_lo >> 8;
71
+ bits[7] = bitlen_lo;
72
+ SHA1_Update(c, bits, 8);
73
+
74
+ for (i = 0; i < 5; i++) {
75
+ uint32_t v = c->hash[i];
76
+ hash[0] = v >> 24;
77
+ hash[1] = v >> 16;
78
+ hash[2] = v >> 8;
79
+ hash[3] = v;
80
+ hash += 4;
81
+ }
82
+}
arm/sha1.h
new
+18
@@ -0,0 +1,18 @@
1
+/*
2
+ * SHA-1 implementation optimized for ARM
3
+ *
4
+ * Copyright: (C) 2005 by Nicolas Pitre <nico@cam.org>
5
+ * Created: September 17, 2005
6
+ */
7
+
8
+#include <stdint.h>
9
+
10
+typedef struct sha_context {
11
+ uint64_t len;
12
+ uint32_t hash[5];
13
+ unsigned char buffer[64];
14
+} SHA_CTX;
15
+
16
+void SHA1_Init(SHA_CTX *c);
17
+void SHA1_Update(SHA_CTX *c, const void *p, unsigned long n);
18
+void SHA1_Final(unsigned char *hash, SHA_CTX *c);
arm/sha1_arm.S
new
+184
@@ -0,0 +1,184 @@
1
+/*
2
+ * SHA transform optimized for ARM
3
+ *
4
+ * Copyright: (C) 2005 by Nicolas Pitre <nico@cam.org>
5
+ * Created: September 17, 2005
6
+ *
7
+ * This program is free software; you can redistribute it and/or modify
8
+ * it under the terms of the GNU General Public License version 2 as
9
+ * published by the Free Software Foundation.
10
+ */
11
+
12
+ .text
13
+ .globl sha_transform
14
+
15
+/*
16
+ * void sha_transform(uint32_t *hash, const unsigned char *data, uint32_t *W);
17
+ *
18
+ * note: the "data" pointer may be unaligned.
19
+ */
20
+
21
+sha_transform:
22
+
23
+ stmfd sp!, {r4 - r8, lr}
24
+
25
+ @ for (i = 0; i < 16; i++)
26
+ @ W[i] = ntohl(((uint32_t *)data)[i]); */
27
+
28
+#ifdef __ARMEB__
29
+ mov r4, r0
30
+ mov r0, r2
31
+ mov r2, #64
32
+ bl memcpy
33
+ mov r2, r0
34
+ mov r0, r4
35
+#else
36
+ mov r3, r2
37
+ mov lr, #16
38
+1: ldrb r4, [r1], #1
39
+ ldrb r5, [r1], #1
40
+ ldrb r6, [r1], #1
41
+ ldrb r7, [r1], #1
42
+ subs lr, lr, #1
43
+ orr r5, r5, r4, lsl #8
44
+ orr r6, r6, r5, lsl #8
45
+ orr r7, r7, r6, lsl #8
46
+ str r7, [r3], #4
47
+ bne 1b
48
+#endif
49
+
50
+ @ for (i = 0; i < 64; i++)
51
+ @ W[i+16] = ror(W[i+13] ^ W[i+8] ^ W[i+2] ^ W[i], 31);
52
+
53
+ sub r3, r2, #4
54
+ mov lr, #64
55
+2: ldr r4, [r3, #4]!
56
+ subs lr, lr, #1
57
+ ldr r5, [r3, #8]
58
+ ldr r6, [r3, #32]
59
+ ldr r7, [r3, #52]
60
+ eor r4, r4, r5
61
+ eor r4, r4, r6
62
+ eor r4, r4, r7
63
+ mov r4, r4, ror #31
64
+ str r4, [r3, #64]
65
+ bne 2b
66
+
67
+ /*
68
+ * The SHA functions are:
69
+ *
70
+ * f1(B,C,D) = (D ^ (B & (C ^ D)))
71
+ * f2(B,C,D) = (B ^ C ^ D)
72
+ * f3(B,C,D) = ((B & C) | (D & (B | C)))
73
+ *
74
+ * Then the sub-blocks are processed as follows:
75
+ *
76
+ * A' = ror(A, 27) + f(B,C,D) + E + K + *W++
77
+ * B' = A
78
+ * C' = ror(B, 2)
79
+ * D' = C
80
+ * E' = D
81
+ *
82
+ * We therefore unroll each loop 5 times to avoid register shuffling.
83
+ * Also the ror for C (and also D and E which are successivelyderived
84
+ * from it) is applied in place to cut on an additional mov insn for
85
+ * each round.
86
+ */
87
+
88
+ .macro sha_f1, A, B, C, D, E
89
+ ldr r3, [r2], #4
90
+ eor ip, \C, \D
91
+ add \E, r1, \E, ror #2
92
+ and ip, \B, ip, ror #2
93
+ add \E, \E, \A, ror #27
94
+ eor ip, ip, \D, ror #2
95
+ add \E, \E, r3
96
+ add \E, \E, ip
97
+ .endm
98
+
99
+ .macro sha_f2, A, B, C, D, E
100
+ ldr r3, [r2], #4
101
+ add \E, r1, \E, ror #2
102
+ eor ip, \B, \C, ror #2
103
+ add \E, \E, \A, ror #27
104
+ eor ip, ip, \D, ror #2
105
+ add \E, \E, r3
106
+ add \E, \E, ip
107
+ .endm
108
+
109
+ .macro sha_f3, A, B, C, D, E
110
+ ldr r3, [r2], #4
111
+ add \E, r1, \E, ror #2
112
+ orr ip, \B, \C, ror #2
113
+ add \E, \E, \A, ror #27
114
+ and ip, ip, \D, ror #2
115
+ add \E, \E, r3
116
+ and r3, \B, \C, ror #2
117
+ orr ip, ip, r3
118
+ add \E, \E, ip
119
+ .endm
120
+
121
+ ldmia r0, {r4 - r8}
122
+
123
+ mov lr, #4
124
+ ldr r1, .L_sha_K + 0
125
+
126
+ /* adjust initial values */
127
+ mov r6, r6, ror #30
128
+ mov r7, r7, ror #30
129
+ mov r8, r8, ror #30
130
+
131
+3: subs lr, lr, #1
132
+ sha_f1 r4, r5, r6, r7, r8
133
+ sha_f1 r8, r4, r5, r6, r7
134
+ sha_f1 r7, r8, r4, r5, r6
135
+ sha_f1 r6, r7, r8, r4, r5
136
+ sha_f1 r5, r6, r7, r8, r4
137
+ bne 3b
138
+
139
+ ldr r1, .L_sha_K + 4
140
+ mov lr, #4
141
+
142
+4: subs lr, lr, #1
143
+ sha_f2 r4, r5, r6, r7, r8
144
+ sha_f2 r8, r4, r5, r6, r7
145
+ sha_f2 r7, r8, r4, r5, r6
146
+ sha_f2 r6, r7, r8, r4, r5
147
+ sha_f2 r5, r6, r7, r8, r4
148
+ bne 4b
149
+
150
+ ldr r1, .L_sha_K + 8
151
+ mov lr, #4
152
+
153
+5: subs lr, lr, #1
154
+ sha_f3 r4, r5, r6, r7, r8
155
+ sha_f3 r8, r4, r5, r6, r7
156
+ sha_f3 r7, r8, r4, r5, r6
157
+ sha_f3 r6, r7, r8, r4, r5
158
+ sha_f3 r5, r6, r7, r8, r4
159
+ bne 5b
160
+
161
+ ldr r1, .L_sha_K + 12
162
+ mov lr, #4
163
+
164
+6: subs lr, lr, #1
165
+ sha_f2 r4, r5, r6, r7, r8
166
+ sha_f2 r8, r4, r5, r6, r7
167
+ sha_f2 r7, r8, r4, r5, r6
168
+ sha_f2 r6, r7, r8, r4, r5
169
+ sha_f2 r5, r6, r7, r8, r4
170
+ bne 6b
171
+
172
+ ldmia r0, {r1, r2, r3, ip, lr}
173
+ add r4, r1, r4
174
+ add r5, r2, r5
175
+ add r6, r3, r6, ror #2
176
+ add r7, ip, r7, ror #2
177
+ add r8, lr, r8, ror #2
178
+ stmia r0, {r4 - r8}
179
+
180
+ ldmfd sp!, {r4 - r8, pc}
181
+
182
+.L_sha_K:
183
+ .word 0x5a827999, 0x6ed9eba1, 0x8f1bbcdc, 0xca62c1d6
184
+