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
|
dnl AMD K6 mpn_mod_34lsub1 -- mpn remainder modulo 2**24-1.
dnl Copyright 2000, 2001, 2002 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 3 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 License
dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/.
include(`../config.m4')
C K6: 2.66 cycles/limb
C mp_limb_t mpn_mod_34lsub1 (mp_srcptr src, mp_size_t size)
C
C An attempt was made to use a loop like
C
C L(top):
C adcl (%edx), %eax
C adcl 4(%edx), %ebx
C adcl 8(%edx), %esi
C leal 12(%edx), %edx
C loop L(top)
C
C with %ecx starting from floor(size/3), but it still measured 2.66 c/l.
C The form used instead can save about 6 cycles by not dividing by 3.
C
C In the code used, putting the "leal"s at the top of the loop is necessary
C for the claimed speed, anywhere else costs an extra cycle per loop.
C Perhaps a tight loop like this needs short decode instructions at the
C branch target, which would explain the leal/loop form above taking 8
C cycles instead of 7 too.
defframe(PARAM_SIZE, 8)
defframe(PARAM_SRC, 4)
dnl re-use parameter space
define(SAVE_EBX, `PARAM_SIZE')
define(SAVE_ESI, `PARAM_SRC')
TEXT
ALIGN(16)
PROLOGUE(mpn_mod_34lsub1)
deflit(`FRAME',0)
movl PARAM_SIZE, %eax
movl PARAM_SRC, %edx
subl $2, %eax
ja L(three_or_more)
Zdisp( movl, 0,(%edx), %eax) C avoid code cache line boundary
jne L(one)
movl %eax, %ecx
movl 4(%edx), %edx
shrl $24, %eax C src[0] high
andl $0x00FFFFFF, %ecx C src[0] low
addl %ecx, %eax
movl %edx, %ecx
shll $8, %edx
andl $0x00FFFF00, %edx C src[1] high
shrl $16, %ecx C src[1] low
addl %ecx, %eax
addl %edx, %eax
L(one):
ret
L(three_or_more):
C eax size-2
C ebx
C ecx
C edx src
movl %ebx, SAVE_EBX
xorl %ebx, %ebx
movl %esi, SAVE_ESI
pushl %edi FRAME_pushl()
xorl %esi, %esi
xorl %edi, %edi C and clear carry flag
L(top):
C eax counter, limbs
C ebx acc 0mod3
C ecx
C edx src, incrementing
C esi acc 1mod3
C edi acc 2mod3
C ebp
leal -2(%eax), %eax
leal 12(%edx), %edx
adcl -12(%edx), %ebx
adcl -8(%edx), %esi
adcl -4(%edx), %edi
decl %eax
jg L(top)
C ecx is -3, -2 or -1 representing 0, 1 or 2 more limbs, respectively
movb $0, %cl
incl %eax
js L(combine) C 0 more
Zdisp( adcl, 0,(%edx), %ebx) C avoid code cache line crossings
movb $8, %cl
decl %eax
js L(combine) C 1 more
adcl 4(%edx), %esi
movb $16, %cl
L(combine):
sbbl %edx, %edx
shll %cl, %edx C carry
movl %ebx, %eax C 0mod3
shrl $24, %eax C 0mod3 high
andl $0x00FFFFFF, %ebx C 0mod3 low
subl %edx, %eax C apply carry
movl %esi, %ecx C 1mod3
shrl $16, %esi C 1mod3 high
addl %ebx, %eax C apply 0mod3 low
andl $0x0000FFFF, %ecx
addl %esi, %eax C apply 1mod3 high
shll $8, %ecx C 1mod3 low
movl %edi, %edx C 2mod3
shrl $8, %edx C 2mod3 high
addl %ecx, %eax C apply 1mod3 low
addl %edx, %eax C apply 2mod3 high
andl $0x000000FF, %edi
shll $16, %edi C 2mod3 low
movl SAVE_EBX, %ebx
addl %edi, %eax C apply 2mod3 low
movl SAVE_ESI, %esi
popl %edi
ret
EPILOGUE()
|
