xref: /netbsd-src/external/lgpl3/gmp/dist/mpn/x86_64/README (revision 7d62b00eb9ad855ffcd7da46b41e23feb5476fac)
1Copyright 2003, 2004, 2006, 2008 Free Software Foundation, Inc.
2
3This file is part of the GNU MP Library.
4
5The GNU MP Library is free software; you can redistribute it and/or modify
6it under the terms of either:
7
8  * the GNU Lesser General Public License as published by the Free
9    Software Foundation; either version 3 of the License, or (at your
10    option) any later version.
11
12or
13
14  * the GNU General Public License as published by the Free Software
15    Foundation; either version 2 of the License, or (at your option) any
16    later version.
17
18or both in parallel, as here.
19
20The GNU MP Library is distributed in the hope that it will be useful, but
21WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
22or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
23for more details.
24
25You should have received copies of the GNU General Public License and the
26GNU Lesser General Public License along with the GNU MP Library.  If not,
27see https://www.gnu.org/licenses/.
28
29
30
31
32
33			AMD64 MPN SUBROUTINES
34
35
36This directory contains mpn functions for AMD64 chips.  It is also useful
37for 64-bit Pentiums, and "Core 2".
38
39
40		     RELEVANT OPTIMIZATION ISSUES
41
42The Opteron and Athlon64 can sustain up to 3 instructions per cycle, but in
43practice that is only possible for integer instructions.  But almost any
44three integer instructions can issue simultaneously, including any 3 ALU
45operations, including shifts.  Up to two memory operations can issue each
46cycle.
47
48Scheduling typically requires that load-use instructions are split into
49separate load and use instructions.  That requires more decode resources,
50and it is rarely a win.  Opteron/Athlon64 have deep out-of-order core.
51
52
53Optimizing for 64-bit Pentium4 is probably a waste of time, as the most
54critical instructions are very poorly implemented here.  Perhaps we could
55save a cycle or two, but the most common loops now run at between 10 and 22
56cycles, so a saved cycle isn't too exciting.
57
58
59The new spin of the venerable P6 core, the "Core 2" is much better than the
60Pentium4 for the GMP loops.  Its integer pipeline is somewhat similar to to
61the Opteron/Athlon64 pipeline, except that the GMP favourites ADC/SBB and
62MUL are slower.  Furthermore, an INC/DEC followed by ADC/SBB incur a
63pipeline stall of around 10 cycles.  The default mpn_add_n and mpn_sub_n
64code suffers badly from the stall.  The code in the core2 subdirectory uses
65the almost forgotten instruction JRCXZ for loop control, and updates the
66induction variable using LEA.
67
68
69
70REFERENCES
71
72"System V Application Binary Interface AMD64 Architecture Processor
73Supplement", draft version 0.99, December 2007.
74http://www.x86-64.org/documentation/abi.pdf
75