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
|
Copyright 2001 Free Software Foundation, Inc.
This file is part of the GNU MP Library.
The GNU MP Library is free software; you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation; either version 2.1 of the License, or (at your
option) any later version.
The GNU MP Library is distributed in the hope that it will be useful, but
WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
License for more details.
You should have received a copy of the GNU Lesser General Public License
along with the GNU MP Library; see the file COPYING.LIB. If not, write to
the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA.
INTEL PENTIUM-4 MPN SUBROUTINES
This directory contains mpn functions optimized for Intel Pentium-4.
The mmx subdirectory has routines using MMX instructions, the sse2
subdirectory has routines using SSE2 instructions. All P4s have these, the
separate directories are just so configure can omit that code if the
assembler doesn't support it.
STATUS
cycles/limb
mpn_add_n/sub_n 4 normal, 6 in-place
mpn_mul_1 4 normal, 6 in-place
mpn_addmul_1 6
mpn_submul_1 7
mpn_mul_basecase 6 cycles/crossproduct (approx)
mpn_sqr_basecase 3.5 cycles/crossproduct (approx)
or 7.0 cycles/triangleproduct (approx)
mpn_l/rshift 1.75
The shifts ought to be able to go at 1.5 c/l, but not much effort has been
applied to them yet.
In-place operations, and all addmul, submul, mul_basecase and sqr_basecase
calls, suffer from pipeline anomalies associated with write combining and
movd reads and writes to the same or nearby locations. The movq
instructions do not trigger the same hardware problems. Unfortunately,
using movq and splitting/combining seems to require too many extra
instructions to help. Perhaps future chip steppings will be better.
NOTES
The Pentium-4 pipeline "Netburst", provides for quite a number of surprises.
Many traditional x86 instructions run very slowly, requiring use of
alterative instructions for acceptable performance.
adcl and sbbl are quite slow at 8 cycles for reg->reg. paddq of 32-bits
within a 64-bit mmx register seems better, though the combination
paddq/psrlq when propagating a carry is still a 4 cycle latency.
incl and decl should be avoided, instead use add $1 and sub $1. Apparently
the carry flag is not separately renamed, so incl and decl depend on all
previous flags-setting instructions.
shll and shrl have a 4 cycle latency, or 8 times the latency of the fastest
integer instructions (addl, subl, orl, andl, and some more). shldl and
shrdl seem to have 13 and 15 cycles latency, respectively. Bizarre.
movq mmx -> mmx does have 6 cycle latency, as noted in the documentation.
pxor/por or similar combination at 2 cycles latency can be used instead.
The movq however executes in the float unit, thereby saving MMX execution
resources. With the right juggling, data moves shouldn't be on a dependent
chain.
L1 is write-through, but the write-combining sounds like it does enough to
not require explicit destination prefetching.
xmm registers so far haven't found a use, but not much effort has been
expended. A configure test for whether the operating system knows
fxsave/fxrestor will be needed if they're used.
REFERENCES
Intel Pentium-4 processor manuals,
http://developer.intel.com/design/pentium4/manuals
"Intel Pentium 4 Processor Optimization Reference Manual", Intel, 2001,
order number 248966. Available on-line:
http://developer.intel.com/design/pentium4/manuals/248966.htm
----------------
Local variables:
mode: text
fill-column: 76
End:
|
