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 K6 MPN SUBROUTINES
36 This directory contains code optimized for AMD K6 CPUs, meaning K6, K6-2 and
39 The mmx subdirectory has MMX code suiting plain K6, the k62mmx subdirectory
40 has MMX code suiting K6-2 and K6-3. All chips in the K6 family have MMX,
41 the separate directories are just so that ./configure can omit them if the
42 assembler doesn't support MMX.
49 Times for the loops, with all code and data in L1 cache, are as follows.
53 mpn_add_n/sub_n 3.25 normal, 2.75 in-place
56 mpn_add/submul_1 7.65-8.4 (varying with data values)
58 mpn_mul_basecase 9.25 cycles/crossproduct (approx)
59 mpn_sqr_basecase 4.7 cycles/crossproduct (approx)
60 or 9.2 cycles/triangleproduct (approx)
72 K6-2 and K6-3 have dual-issue MMX and get the following improvements.
77 Prefetching of sources hasn't yet given any joy. With the 3DNow "prefetch"
78 instruction, code seems to run slower, and with just "mov" loads it doesn't
79 seem faster. Results so far are inconsistent. The K6 does a hardware
80 prefetch of the second cache line in a sector, so the penalty for not
81 prefetching in software is reduced.
88 All K6 family chips have MMX, but only K6-2 and K6-3 have 3DNow.
90 Plain K6 executes MMX instructions only in the X pipe, but K6-2 and K6-3 can
91 execute them in both X and Y (and in both together).
93 Branch misprediction penalty is 1 to 4 cycles (Optimization Manual
96 Write-allocate L1 data cache means prefetching of destinations is unnecessary.
97 Store queue is 7 entries of 64 bits each.
99 Floating point multiplications can be done in parallel with integer
100 multiplications, but there doesn't seem to be any way to make use of this.
106 Unrolled loops are used to reduce looping overhead. The unrolling is
107 configurable up to 32 limbs/loop for most routines, up to 64 for some.
109 Sometimes computed jumps into the unrolling are used to handle sizes not a
110 multiple of the unrolling. An attractive feature of this is that times
111 smoothly increase with operand size, but an indirect jump is about 6 cycles
112 and the setups about another 6, so it depends on how much the unrolled code
113 is faster than a simple loop as to whether a computed jump ought to be used.
115 Position independent code is implemented using a call to get eip for
116 computed jumps and a ret is always done, rather than an addl $4,%esp or a
117 popl, so the CPU return address branch prediction stack stays synchronised
118 with the actual stack in memory. Such a call however still costs 4 to 7
121 Branch prediction, in absence of any history, will guess forward jumps are
122 not taken and backward jumps are taken. Where possible it's arranged that
123 the less likely or less important case is under a taken forward jump.
129 Putting emms or femms as late as possible in a routine seems to be fastest.
130 Perhaps an emms or femms stalls until all outstanding MMX instructions have
131 completed, so putting it later gives them a chance to complete on their own,
132 in parallel with other operations (like register popping).
134 The Optimization Manual chapter 5 recommends using a femms on K6-2 and K6-3
135 at the start of a routine, in case it's been preceded by x87 floating point
136 operations. This isn't done because in gmp programs it's expected that x87
137 floating point won't be much used and that chances are an mpn routine won't
138 have been preceded by any x87 code.
144 Instructions in general code are shown paired if they can decode and execute
145 together, meaning two short decode instructions with the second not
146 depending on the first, only the first using the shifter, no more than one
147 load, and no more than one store.
149 K6 does some out of order execution so the pairings aren't essential, they
150 just show what slots might be available. When decoding is the limiting
151 factor things can be scheduled that might not execute until later.
159 - if an opcode/modrm or 0Fh/opcode/modrm crosses a cache line boundary,
160 short decode is inhibited. The cross.pl script detects this.
162 - loops and branch targets should be aligned to 16 bytes, or ensure at least
163 2 instructions before a 32 byte boundary. This makes use of the 16 byte
168 - (%esi) degrades decoding from short to vector. 0(%esi) doesn't have this
169 problem, and can be used as an equivalent, or easier is just to use a
170 different register, like %ebx.
172 - K6 and pre-CXT core K6-2 have the following problem. (K6-2 CXT and K6-3
173 have it fixed, these being cpuid function 1 signatures 0x588 to 0x58F).
175 If more than 3 bytes are needed to determine instruction length then
176 decoding degrades from direct to long, or from long to vector. This
177 happens with forms like "0F opcode mod/rm" with mod/rm=00-xxx-100 since
178 with mod=00 the sib determines whether there's a displacement.
180 This affects all MMX and 3DNow instructions, and others with an 0F prefix,
181 like movzbl. The modes affected are anything with an index and no
182 displacement, or an index but no base, and this includes (%esp) which is
185 The cross.pl script detects problem cases. The workaround is to always
186 use a displacement, and to do this with Zdisp if it's zero so the
187 assembler doesn't discard it.
189 See Optimization Manual rev D page 67 and 3DNow Porting Guide rev B pages
194 - indirect jumps and calls are not branch predicted, they measure about 6
199 - adcl 2 cycles of decode, maybe 2 cycles executing in the X pipe
203 - jecxz 2 cycles taken, 13 not taken (optimization manual says 7 not taken)
204 - divl 20 cycles back-to-back
205 - imull 2 decode, 3 execute
206 - mull 2 decode, 3 execute (optimization manual decoding sample)
208 - rcll/rcrl implicit by one bit: 2 cycles
209 immediate or %cl count: 11 + 2 per bit for dword
210 13 + 4 per bit for byte
212 - xchgl %eax,reg 1.5 cycles, back-to-back (strange)
213 reg,reg 2 cycles, back-to-back
220 "AMD-K6 Processor Code Optimization Application Note", AMD publication
221 number 21924, revision D amendment 0, January 2000. This describes K6-2 and
222 K6-3. Available on-line,
224 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21924.pdf
226 "AMD-K6 MMX Enhanced Processor x86 Code Optimization Application Note", AMD
227 publication number 21828, revision A amendment 0, August 1997. This is an
228 older edition of the above document, describing plain K6. Available
231 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21828.pdf
233 "3DNow Technology Manual", AMD publication number 21928G/0-March 2000.
234 This describes the femms and prefetch instructions, but nothing else from
235 3DNow has been used. Available on-line,
237 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/21928.pdf
239 "3DNow Instruction Porting Guide", AMD publication number 22621, revision B,
240 August 1999. This has some notes on general K6 optimizations as well as
241 3DNow. Available on-line,
243 http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22621.pdf