Copyright 1996, 1999-2001, 2003 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 either:
* the GNU Lesser General Public License as published by the Free
Software Foundation; either version 3 of the License, or (at your
option) any later version.
or
* the GNU General Public License as published by the Free Software
Foundation; either version 2 of the License, or (at your option) any
later version.
or both in parallel, as here.
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 General Public License
for more details.
You should have received copies of the GNU General Public License and the
GNU Lesser General Public License along with the GNU MP Library. If not,
see https://www.gnu.org/licenses/.
INTEL PENTIUM P5 MPN SUBROUTINES
This directory contains mpn functions optimized for Intel Pentium (P5,P54)
processors. The mmx subdirectory has additional code for Pentium with MMX
(P55).
STATUS
cycles/limb
mpn_add_n/sub_n 2.375
mpn_mul_1 12.0
mpn_add/submul_1 14.0
mpn_mul_basecase 14.2 cycles/crossproduct (approx)
mpn_sqr_basecase 8 cycles/crossproduct (approx)
or 15.5 cycles/triangleproduct (approx)
mpn_l/rshift 5.375 normal (6.0 on P54)
1.875 special shift by 1 bit
mpn_divrem_1 44.0
mpn_mod_1 28.0
mpn_divexact_by3 15.0
mpn_copyi/copyd 1.0
Pentium MMX gets the following improvements
mpn_l/rshift 1.75
mpn_mul_1 12.0 normal, 7.0 for 16-bit multiplier
mpn_add_n and mpn_sub_n run at asymptotically 2 cycles/limb. Due to loop
overhead and other delays (cache refill?), they run at or near 2.5
cycles/limb.
mpn_mul_1, mpn_addmul_1, mpn_submul_1 all run 1 cycle faster than they
should. Intel documentation says a mul instruction is 10 cycles, but it
measures 9 and the routines using it run as 9.
P55 MMX AND X87
The cost of switching between MMX and x87 floating point on P55 is about 100
cycles (fld1/por/emms for instance). In order to avoid that the two aren't
mixed and currently that means using MMX and not x87.
MMX offers a big speedup for lshift and rshift, and a nice speedup for
16-bit multipliers in mpn_mul_1. If fast code using x87 is found then
perhaps the preference for MMX will be reversed.
P54 SHLDL
mpn_lshift and mpn_rshift run at about 6 cycles/limb on P5 and P54, but the
documentation indicates that they should take only 43/8 = 5.375 cycles/limb,
or 5 cycles/limb asymptotically. The P55 runs them at the expected speed.
It seems that on P54 a shldl or shrdl allows pairing in one following cycle,
but not two. For example, back to back repetitions of the following
shldl( %cl, %eax, %ebx)
xorl %edx, %edx
xorl %esi, %esi
run at 5 cycles, as expected, but repetitions of the following run at 7
cycles, whereas 6 would be expected (and is achieved on P55),
shldl( %cl, %eax, %ebx)
xorl %edx, %edx
xorl %esi, %esi
xorl %edi, %edi
xorl %ebp, %ebp
Three xorls run at 7 cycles too, so it doesn't seem to be just that pairing
inhibited is only in the second following cycle (or something like that).
Avoiding this problem would bring P54 shifts down from 6.0 c/l to 5.5 with a
pattern of shift, 2 loads, shift, 2 stores, shift, etc. A start has been
made on something like that, but it's not yet complete.
OTHER NOTES
Prefetching Destinations
Pentium doesn't allocate cache lines on writes, unlike most other modern
processors. Since the functions in the mpn class do array writes, we
have to handle allocating the destination cache lines by reading a word
from it in the loops, to achieve the best performance.
Prefetching Sources
Prefetching of sources is pointless since there's no out-of-order loads.
Any load instruction blocks until the line is brought to L1, so it may
as well be the load that wants the data which blocks.
Data Cache Bank Clashes
Pairing of memory operations requires that the two issued operations
refer to different cache banks (ie. different addresses modulo 32
bytes). The simplest way to ensure this is to read/write two words from
the same object. If we make operations on different objects, they might
or might not be to the same cache bank.
PIC %eip Fetching
A simple call $+5 and popl can be used to get %eip, there's no need to
balance calls and returns since P5 doesn't have any return stack branch
prediction.
Float Multiplies
fmul is pairable and can be issued every 2 cycles (with a 4 cycle
latency for data ready to use). This is a lot better than integer mull
or imull at 9 cycles non-pairing. Unfortunately the advantage is
quickly eaten away by needing to throw data through memory back to the
integer registers to adjust for fild and fist being signed, and to do
things like propagating carry bits.
REFERENCES
"Intel Architecture Optimization Manual", 1997, order number 242816. This
is mostly about P5, the parts about P6 aren't relevant. Available on-line:
http://download.intel.com/design/PentiumII/manuals/242816.htm
----------------
Local variables:
mode: text
fill-column: 76
End: