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
|
dnl AMD K7 mpn_mul_1 -- mpn by limb multiply.
dnl
dnl K7: 3.4 cycles/limb (at 16 limbs/loop).
dnl Copyright (C) 1999, 2000 Free Software Foundation, Inc.
dnl
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
dnl modify it under the terms of the GNU Lesser General Public License as
dnl published by the Free Software Foundation; either version 2.1 of the
dnl License, or (at your option) any later version.
dnl
dnl The GNU MP Library is distributed in the hope that it will be useful,
dnl but WITHOUT ANY WARRANTY; without even the implied warranty of
dnl MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
dnl Lesser General Public License for more details.
dnl
dnl You should have received a copy of the GNU Lesser General Public
dnl License along with the GNU MP Library; see the file COPYING.LIB. If
dnl not, write to the Free Software Foundation, Inc., 59 Temple Place -
dnl Suite 330, Boston, MA 02111-1307, USA.
include(`../config.m4')
dnl K7: UNROLL_COUNT cycles/limb
dnl 8 3.9
dnl 16 3.4
dnl 32 3.4
dnl 64 3.35
dnl Maximum possible with the current code is 64.
deflit(UNROLL_COUNT, 16)
C mp_limb_t mpn_mul_1 (mp_ptr dst, mp_srcptr src, mp_size_t size,
C mp_limb_t multiplier);
C mp_limb_t mpn_mul_1c (mp_ptr dst, mp_srcptr src, mp_size_t size,
C mp_limb_t multiplier, mp_limb_t carry);
C
C Multiply src,size by mult and store the result in dst,size.
C Return the carry limb from the top of the result.
C
C mpn_mul_1c() accepts an initial carry for the calculation, it's added into
C the low limb of the destination.
C
C Variations on the unrolled loop have been tried, with the current
C registers or with the counter on the stack to free up ecx. The current
C code is the fastest found.
C
C An interesting effect is that removing the stores "movl %ebx, disp0(%edi)"
C from the unrolled loop actually slows it down to 5.0 cycles/limb. Code
C with this change can be tested on sizes of the form UNROLL_COUNT*n+1
C without having to change the computed jump. There's obviously something
C fishy going on, perhaps with what execution units the mul needs.
defframe(PARAM_CARRY, 20)
defframe(PARAM_MULTIPLIER,16)
defframe(PARAM_SIZE, 12)
defframe(PARAM_SRC, 8)
defframe(PARAM_DST, 4)
defframe(SAVE_EBP, -4)
defframe(SAVE_EDI, -8)
defframe(SAVE_ESI, -12)
defframe(SAVE_EBX, -16)
deflit(STACK_SPACE, 16)
dnl Must have UNROLL_THRESHOLD >= 2, since the unrolled loop can't handle 1.
ifdef(`PIC',`
deflit(UNROLL_THRESHOLD, 7)
',`
deflit(UNROLL_THRESHOLD, 5)
')
.text
ALIGN(32)
PROLOGUE(mpn_mul_1c)
deflit(`FRAME',0)
movl PARAM_CARRY, %edx
jmp LF(mpn_mul_1,start_nc)
EPILOGUE()
PROLOGUE(mpn_mul_1)
deflit(`FRAME',0)
xorl %edx, %edx C initial carry
L(start_nc):
movl PARAM_SIZE, %ecx
subl $STACK_SPACE, %esp
deflit(`FRAME', STACK_SPACE)
movl %edi, SAVE_EDI
movl %ebx, SAVE_EBX
movl %edx, %ebx
movl %esi, SAVE_ESI
movl PARAM_SRC, %esi
cmpl $UNROLL_THRESHOLD, %ecx
movl PARAM_DST, %edi
movl %ebp, SAVE_EBP
jae L(unroll)
leal (%esi,%ecx,4), %esi
leal (%edi,%ecx,4), %edi
negl %ecx
movl PARAM_MULTIPLIER, %ebp
L(simple):
C eax scratch
C ebx carry
C ecx counter (negative)
C edx scratch
C esi src
C edi dst
C ebp multiplier
movl (%esi,%ecx,4), %eax
mull %ebp
addl %ebx, %eax
movl %eax, (%edi,%ecx,4)
movl $0, %ebx
adcl %edx, %ebx
incl %ecx
jnz L(simple)
movl %ebx, %eax
movl SAVE_EBX, %ebx
movl SAVE_ESI, %esi
movl SAVE_EDI, %edi
movl SAVE_EBP, %ebp
addl $STACK_SPACE, %esp
ret
C -----------------------------------------------------------------------------
C The mov to load the next source limb is done well ahead of the mul, this
C is necessary for full speed. It leads to one limb handled separately
C after the loop.
C
C When unrolling to 32 or more, an offset of +4 is used on the src pointer,
C to avoid having an 0x80 displacement in the code for the last limb in the
C unrolled loop. This is for a fair comparison between 16 and 32 unrolling.
ifelse(eval(UNROLL_COUNT >= 32),1,`
deflit(SRC_OFFSET,4)
',`
deflit(SRC_OFFSET,)
')
C this is offset 0x62, so close enough to aligned
L(unroll):
C eax
C ebx initial carry
C ecx size
C edx
C esi src
C edi dst
C ebp
deflit(`FRAME', STACK_SPACE)
leal -1(%ecx), %edx C one limb handled at end
leal -2(%ecx), %ecx C and ecx is one less than edx
movl %ebp, SAVE_EBP
negl %edx
shrl $UNROLL_LOG2, %ecx C unrolled loop counter
movl (%esi), %eax C src low limb
andl $UNROLL_MASK, %edx
movl PARAM_DST, %edi
movl %edx, %ebp
shll $4, %edx
C 17 code bytes per limb
ifdef(`PIC',`
call L(add_eip_to_edx)
L(here):
',`
leal L(entry) (%edx,%ebp), %edx
')
negl %ebp
leal ifelse(UNROLL_BYTES,256,128+) SRC_OFFSET(%esi,%ebp,4), %esi
leal ifelse(UNROLL_BYTES,256,128) (%edi,%ebp,4), %edi
movl PARAM_MULTIPLIER, %ebp
jmp *%edx
ifdef(`PIC',`
L(add_eip_to_edx):
C See README.family about old gas bugs
leal (%edx,%ebp), %edx
addl $L(entry)-L(here), %edx
addl (%esp), %edx
ret
')
C ----------------------------------------------------------------------------
ALIGN(32)
L(top):
C eax next src limb
C ebx carry
C ecx counter
C edx scratch
C esi src+4
C edi dst
C ebp multiplier
C
C 17 code bytes per limb processed
L(entry):
forloop(i, 0, UNROLL_COUNT-1, `
deflit(`disp_dst', eval(i*4 ifelse(UNROLL_BYTES,256,-128)))
deflit(`disp_src', eval(disp_dst + 4-(SRC_OFFSET-0)))
mull %ebp
addl %eax, %ebx
Zdisp( movl, disp_src,(%esi), %eax)
Zdisp( movl, %ebx, disp_dst,(%edi))
movl $0, %ebx
adcl %edx, %ebx
')
decl %ecx
leal UNROLL_BYTES(%esi), %esi
leal UNROLL_BYTES(%edi), %edi
jns L(top)
deflit(`disp0', ifelse(UNROLL_BYTES,256,-128))
mull %ebp
addl %eax, %ebx
movl $0, %eax
movl SAVE_ESI, %esi
movl %ebx, disp0(%edi)
movl SAVE_EBX, %ebx
movl SAVE_EDI, %edi
adcl %edx, %eax
movl SAVE_EBP, %ebp
addl $STACK_SPACE, %esp
ret
EPILOGUE()
|