1 Copyright 1996, 1999, 2000, 2001, 2003 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 the GNU Lesser General Public License as published by
7 the Free Software Foundation; either version 3 of the License, or (at your
8 option) any later version.
10 The GNU MP Library is distributed in the hope that it will be useful, but
11 WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
12 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public
13 License for more details.
15 You should have received a copy of the GNU Lesser General Public License
16 along with the GNU MP Library. If not, see http://www.gnu.org/licenses/.
22 INTEL PENTIUM P5 MPN SUBROUTINES
25 This directory contains mpn functions optimized for Intel Pentium (P5,P54)
26 processors. The mmx subdirectory has additional code for Pentium with MMX
39 mpn_mul_basecase 14.2 cycles/crossproduct (approx)
41 mpn_sqr_basecase 8 cycles/crossproduct (approx)
42 or 15.5 cycles/triangleproduct (approx)
44 mpn_l/rshift 5.375 normal (6.0 on P54)
45 1.875 special shift by 1 bit
53 Pentium MMX gets the following improvements
57 mpn_mul_1 12.0 normal, 7.0 for 16-bit multiplier
60 mpn_add_n and mpn_sub_n run at asymptotically 2 cycles/limb. Due to loop
61 overhead and other delays (cache refill?), they run at or near 2.5
64 mpn_mul_1, mpn_addmul_1, mpn_submul_1 all run 1 cycle faster than they
65 should. Intel documentation says a mul instruction is 10 cycles, but it
66 measures 9 and the routines using it run as 9.
72 The cost of switching between MMX and x87 floating point on P55 is about 100
73 cycles (fld1/por/emms for instance). In order to avoid that the two aren't
74 mixed and currently that means using MMX and not x87.
76 MMX offers a big speedup for lshift and rshift, and a nice speedup for
77 16-bit multipliers in mpn_mul_1. If fast code using x87 is found then
78 perhaps the preference for MMX will be reversed.
85 mpn_lshift and mpn_rshift run at about 6 cycles/limb on P5 and P54, but the
86 documentation indicates that they should take only 43/8 = 5.375 cycles/limb,
87 or 5 cycles/limb asymptotically. The P55 runs them at the expected speed.
89 It seems that on P54 a shldl or shrdl allows pairing in one following cycle,
90 but not two. For example, back to back repetitions of the following
92 shldl( %cl, %eax, %ebx)
96 run at 5 cycles, as expected, but repetitions of the following run at 7
97 cycles, whereas 6 would be expected (and is achieved on P55),
99 shldl( %cl, %eax, %ebx)
105 Three xorls run at 7 cycles too, so it doesn't seem to be just that pairing
106 inhibited is only in the second following cycle (or something like that).
108 Avoiding this problem would bring P54 shifts down from 6.0 c/l to 5.5 with a
109 pattern of shift, 2 loads, shift, 2 stores, shift, etc. A start has been
110 made on something like that, but it's not yet complete.
117 Prefetching Destinations
119 Pentium doesn't allocate cache lines on writes, unlike most other modern
120 processors. Since the functions in the mpn class do array writes, we
121 have to handle allocating the destination cache lines by reading a word
122 from it in the loops, to achieve the best performance.
126 Prefetching of sources is pointless since there's no out-of-order loads.
127 Any load instruction blocks until the line is brought to L1, so it may
128 as well be the load that wants the data which blocks.
130 Data Cache Bank Clashes
132 Pairing of memory operations requires that the two issued operations
133 refer to different cache banks (ie. different addresses modulo 32
134 bytes). The simplest way to ensure this is to read/write two words from
135 the same object. If we make operations on different objects, they might
136 or might not be to the same cache bank.
140 A simple call $+5 and popl can be used to get %eip, there's no need to
141 balance calls and returns since P5 doesn't have any return stack branch
146 fmul is pairable and can be issued every 2 cycles (with a 4 cycle
147 latency for data ready to use). This is a lot better than integer mull
148 or imull at 9 cycles non-pairing. Unfortunately the advantage is
149 quickly eaten away by needing to throw data through memory back to the
150 integer registers to adjust for fild and fist being signed, and to do
151 things like propagating carry bits.
159 "Intel Architecture Optimization Manual", 1997, order number 242816. This
160 is mostly about P5, the parts about P6 aren't relevant. Available on-line:
162 http://download.intel.com/design/PentiumII/manuals/242816.htm