1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
|
dnl ARM Neon mpn_lshift and mpn_rshift.
dnl Contributed to the GNU project by Torbjörn Granlund.
dnl Copyright 2013 Free Software Foundation, Inc.
dnl This file is part of the GNU MP Library.
dnl
dnl The GNU MP Library is free software; you can redistribute it and/or modify
dnl it under the terms of either:
dnl
dnl * the GNU Lesser General Public License as published by the Free
dnl Software Foundation; either version 3 of the License, or (at your
dnl option) any later version.
dnl
dnl or
dnl
dnl * the GNU General Public License as published by the Free Software
dnl Foundation; either version 2 of the License, or (at your option) any
dnl later version.
dnl
dnl or both in parallel, as here.
dnl
dnl The GNU MP Library is distributed in the hope that it will be useful, but
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
dnl for more details.
dnl
dnl You should have received copies of the GNU General Public License and the
dnl GNU Lesser General Public License along with the GNU MP Library. If not,
dnl see https://www.gnu.org/licenses/.
include(`../config.m4')
C cycles/limb cycles/limb cycles/limb good
C aligned unaligned best seen for cpu?
C StrongARM - -
C XScale - -
C Cortex-A7 ? ?
C Cortex-A8 ? ?
C Cortex-A9 3 3 Y
C Cortex-A15 1.5 1.5 Y
C We read 64 bits at a time at 32-bit aligned addresses, and except for the
C first and last store, we write using 64-bit aligned addresses. All shifting
C is done on 64-bit words in 'extension' registers.
C
C It should be possible to read also using 64-bit alignment, by manipulating
C the shift count for unaligned operands. Not done, since it does not seem to
C matter for A9 or A15.
C
C This will not work in big-endian mode.
C TODO
C * Try using 128-bit operations. Note that Neon lacks pure 128-bit shifts,
C which might make it tricky.
C * Clean up and simplify.
C * Consider sharing most of the code for lshift and rshift, since the feed-in
C code, the loop, and most of the wind-down code are identical.
C * Replace the basecase code with code using 'extension' registers.
C * Optimise. It is not clear that this loop insn permutation is optimal for
C either A9 or A15.
C INPUT PARAMETERS
define(`rp', `r0')
define(`ap', `r1')
define(`n', `r2')
define(`cnt', `r3')
ifdef(`OPERATION_lshift',`
define(`IFLSH', `$1')
define(`IFRSH', `')
define(`X',`0')
define(`Y',`1')
define(`func',`mpn_lshift')
')
ifdef(`OPERATION_rshift',`
define(`IFLSH', `')
define(`IFRSH', `$1')
define(`X',`1')
define(`Y',`0')
define(`func',`mpn_rshift')
')
MULFUNC_PROLOGUE(mpn_lshift mpn_rshift)
ASM_START(neon)
TEXT
ALIGN(64)
PROLOGUE(func)
IFLSH(` mov r12, n, lsl #2 ')
IFLSH(` add rp, rp, r12 ')
IFLSH(` add ap, ap, r12 ')
cmp n, #4 C SIMD code n limit
ble L(base)
ifdef(`OPERATION_lshift',`
vdup.32 d6, r3 C left shift count is positive
sub r3, r3, #64 C right shift count is negative
vdup.32 d7, r3
mov r12, #-8') C lshift pointer update offset
ifdef(`OPERATION_rshift',`
rsb r3, r3, #0 C right shift count is negative
vdup.32 d6, r3
add r3, r3, #64 C left shift count is positive
vdup.32 d7, r3
mov r12, #8') C rshift pointer update offset
IFLSH(` sub ap, ap, #8 ')
vld1.32 {d19}, [ap], r12 C load initial 2 limbs
vshl.u64 d18, d19, d7 C retval
tst rp, #4 C is rp 64-bit aligned already?
beq L(rp_aligned) C yes, skip
IFLSH(` add ap, ap, #4 ') C move back ap pointer
IFRSH(` sub ap, ap, #4 ') C move back ap pointer
vshl.u64 d4, d19, d6
sub n, n, #1 C first limb handled
IFLSH(` sub rp, rp, #4 ')
vst1.32 {d4[Y]}, [rp]IFRSH(!) C store first limb, rp gets aligned
vld1.32 {d19}, [ap], r12 C load ap[1] and ap[2]
L(rp_aligned):
IFLSH(` sub rp, rp, #8 ')
subs n, n, #6
blt L(two_or_three_more)
tst n, #2
beq L(2)
L(1): vld1.32 {d17}, [ap], r12
vshl.u64 d5, d19, d6
vld1.32 {d16}, [ap], r12
vshl.u64 d0, d17, d7
vshl.u64 d4, d17, d6
sub n, n, #2
b L(mid)
L(2): vld1.32 {d16}, [ap], r12
vshl.u64 d4, d19, d6
vld1.32 {d17}, [ap], r12
vshl.u64 d1, d16, d7
vshl.u64 d5, d16, d6
subs n, n, #4
blt L(end)
L(top): vld1.32 {d16}, [ap], r12
vorr d2, d4, d1
vshl.u64 d0, d17, d7
vshl.u64 d4, d17, d6
vst1.32 {d2}, [rp:64], r12
L(mid): vld1.32 {d17}, [ap], r12
vorr d3, d5, d0
vshl.u64 d1, d16, d7
vshl.u64 d5, d16, d6
vst1.32 {d3}, [rp:64], r12
subs n, n, #4
bge L(top)
L(end): tst n, #1
beq L(evn)
vorr d2, d4, d1
vst1.32 {d2}, [rp:64], r12
b L(cj1)
L(evn): vorr d2, d4, d1
vshl.u64 d0, d17, d7
vshl.u64 d16, d17, d6
vst1.32 {d2}, [rp:64], r12
vorr d2, d5, d0
b L(cj2)
C Load last 2 - 3 limbs, store last 4 - 5 limbs
L(two_or_three_more):
tst n, #1
beq L(l2)
L(l3): vshl.u64 d5, d19, d6
vld1.32 {d17}, [ap], r12
L(cj1): veor d16, d16, d16
IFLSH(` add ap, ap, #4 ')
vld1.32 {d16[Y]}, [ap], r12
vshl.u64 d0, d17, d7
vshl.u64 d4, d17, d6
vorr d3, d5, d0
vshl.u64 d1, d16, d7
vshl.u64 d5, d16, d6
vst1.32 {d3}, [rp:64], r12
vorr d2, d4, d1
vst1.32 {d2}, [rp:64], r12
IFLSH(` add rp, rp, #4 ')
vst1.32 {d5[Y]}, [rp]
vmov.32 r0, d18[X]
bx lr
L(l2): vld1.32 {d16}, [ap], r12
vshl.u64 d4, d19, d6
vshl.u64 d1, d16, d7
vshl.u64 d16, d16, d6
vorr d2, d4, d1
L(cj2): vst1.32 {d2}, [rp:64], r12
vst1.32 {d16}, [rp]
vmov.32 r0, d18[X]
bx lr
define(`tnc', `r12')
L(base):
push {r4, r6, r7, r8}
ifdef(`OPERATION_lshift',`
ldr r4, [ap, #-4]!
rsb tnc, cnt, #32
mov r7, r4, lsl cnt
tst n, #1
beq L(ev) C n even
L(od): subs n, n, #2
bcc L(ed1) C n = 1
ldr r8, [ap, #-4]!
b L(md) C n = 3
L(ev): ldr r6, [ap, #-4]!
subs n, n, #2
beq L(ed) C n = 3
C n = 4
L(tp): ldr r8, [ap, #-4]!
orr r7, r7, r6, lsr tnc
str r7, [rp, #-4]!
mov r7, r6, lsl cnt
L(md): ldr r6, [ap, #-4]!
orr r7, r7, r8, lsr tnc
str r7, [rp, #-4]!
mov r7, r8, lsl cnt
L(ed): orr r7, r7, r6, lsr tnc
str r7, [rp, #-4]!
mov r7, r6, lsl cnt
L(ed1): str r7, [rp, #-4]
mov r0, r4, lsr tnc
')
ifdef(`OPERATION_rshift',`
ldr r4, [ap]
rsb tnc, cnt, #32
mov r7, r4, lsr cnt
tst n, #1
beq L(ev) C n even
L(od): subs n, n, #2
bcc L(ed1) C n = 1
ldr r8, [ap, #4]!
b L(md) C n = 3
L(ev): ldr r6, [ap, #4]!
subs n, n, #2
beq L(ed) C n = 2
C n = 4
L(tp): ldr r8, [ap, #4]!
orr r7, r7, r6, lsl tnc
str r7, [rp], #4
mov r7, r6, lsr cnt
L(md): ldr r6, [ap, #4]!
orr r7, r7, r8, lsl tnc
str r7, [rp], #4
mov r7, r8, lsr cnt
L(ed): orr r7, r7, r6, lsl tnc
str r7, [rp], #4
mov r7, r6, lsr cnt
L(ed1): str r7, [rp], #4
mov r0, r4, lsl tnc
')
pop {r4, r6, r7, r8}
bx r14
EPILOGUE()
|
