1 Copyright 2000, 2001 Free Software Foundation, Inc.
3 This file is part of the GNU MP Library.
5 The GNU MP Library is free software; you can redistribute it and/or modify
6 it under the terms of either:
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.
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
18 or both in parallel, as here.
20 The GNU MP Library is distributed in the hope that it will be useful, but
21 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
22 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
25 You should have received copies of the GNU General Public License and the
26 GNU Lesser General Public License along with the GNU MP Library. If not,
27 see https://www.gnu.org/licenses/.
32 AMD K7 MPN SUBROUTINES
35 This directory contains code optimized for the AMD Athlon CPU.
37 The mmx subdirectory has routines using MMX instructions. All Athlons have
38 MMX, the separate directory is just so that configure can omit it if the
39 assembler doesn't support MMX.
45 Times for the loops, with all code and data in L1 cache.
50 mpn_copyi 0.75 or 1.0 \ varying with data alignment
51 mpn_copyd 0.75 or 1.0 /
53 mpn_divrem_1 17.0 integer part, 15.0 fractional part
60 mpn_addmul/submul_1 3.9
62 mpn_mul_basecase 4.42 cycles/crossproduct (approx)
63 mpn_sqr_basecase 2.3 cycles/crossproduct (approx)
64 or 4.55 cycles/triangleproduct (approx)
66 Prefetching of sources hasn't yet been tried.
72 cmov, MMX, 3DNow and some extensions to MMX and 3DNow are available.
74 Write-allocate L1 data cache means prefetching of destinations is unnecessary.
76 Floating point multiplications can be done in parallel with integer
77 multiplications, but there doesn't seem to be any way to make use of this.
79 Unsigned "mul"s can be issued every 3 cycles. This suggests 3 is a limit on
80 the speed of the multiplication routines. The documentation shows mul
81 executing in IEU0 (or maybe in IEU0 and IEU1 together), so it might be that,
82 to get near 3 cycles code has to be arranged so that nothing else is issued
83 to IEU0. A busy IEU0 could explain why some code takes 4 cycles and other
84 apparently equivalent code takes 5.
90 Unrolled loops are used to reduce looping overhead. The unrolling is
91 configurable up to 32 limbs/loop for most routines and up to 64 for some.
92 The K7 has 64k L1 code cache so quite big unrolling is allowable.
94 Computed jumps into the unrolling are used to handle sizes not a multiple of
95 the unrolling. An attractive feature of this is that times increase
96 smoothly with operand size, but it may be that some routines should just
97 have simple loops to finish up, especially when PIC adds between 2 and 16
100 Position independent code is implemented using a call to get %eip for the
101 computed jumps and a ret is always done, rather than an addl $4,%esp or a
102 popl, so the CPU return address branch prediction stack stays synchronised
103 with the actual stack in memory.
105 Branch prediction, in absence of any history, will guess forward jumps are
106 not taken and backward jumps are taken. Where possible it's arranged that
107 the less likely or less important case is under a taken forward jump.
113 Instructions in general code have been shown grouped if they can execute
114 together, which means up to three direct-path instructions which have no
115 successive dependencies. K7 always decodes three and has out-of-order
116 execution, but the groupings show what slots might be available and what
117 dependency chains exist.
119 When there's vector-path instructions an effort is made to get triplets of
120 direct-path instructions in between them, even if there's dependencies,
121 since this maximizes decoding throughput and might save a cycle or two if
122 decoding is the limiting factor.
129 divl 39 cycles back-to-back
131 loop 1 cycle vector (decl/jnz opens up one decode slot)
134 mull issue every 3 cycles, latency 4 cycles low word, 6 cycles high word
135 popl vector (use movl for more than one pop)
136 pushl direct, will pair with a load
137 shrdl %cl vector, 3 cycles, seems to be 3 decode too
138 xorl r,r false read dependency recognised
144 "AMD Athlon Processor X86 Code Optimization Guide", AMD publication number
145 22007, revision K, February 2002. Available on-line,
147 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf
149 "3DNow Technology Manual", AMD publication number 21928G/0-March 2000.
150 This describes the femms and prefetch instructions. Available on-line,
152 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21928.pdf
154 "AMD Extensions to the 3DNow and MMX Instruction Sets Manual", AMD
155 publication number 22466, revision D, March 2000. This describes
156 instructions added in the Athlon processor, such as pswapd and the extra
157 prefetch forms. Available on-line,
159 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22466.pdf
161 "3DNow Instruction Porting Guide", AMD publication number 22621, revision B,
162 August 1999. This has some notes on general Athlon optimizations as well as
163 3DNow. Available on-line,
165 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22621.pdf