Copyright 2001 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-4 MPN SUBROUTINES This directory contains mpn functions optimized for Intel Pentium-4. The mmx subdirectory has routines using MMX instructions, the sse2 subdirectory has routines using SSE2 instructions. All P4s have these, the separate directories are just so configure can omit that code if the assembler doesn't support it. STATUS cycles/limb mpn_add_n/sub_n 4 normal, 6 in-place mpn_mul_1 4 normal, 6 in-place mpn_addmul_1 6 mpn_submul_1 7 mpn_mul_basecase 6 cycles/crossproduct (approx) mpn_sqr_basecase 3.5 cycles/crossproduct (approx) or 7.0 cycles/triangleproduct (approx) mpn_l/rshift 1.75 The shifts ought to be able to go at 1.5 c/l, but not much effort has been applied to them yet. In-place operations, and all addmul, submul, mul_basecase and sqr_basecase calls, suffer from pipeline anomalies associated with write combining and movd reads and writes to the same or nearby locations. The movq instructions do not trigger the same hardware problems. Unfortunately, using movq and splitting/combining seems to require too many extra instructions to help. Perhaps future chip steppings will be better. NOTES The Pentium-4 pipeline "Netburst", provides for quite a number of surprises. Many traditional x86 instructions run very slowly, requiring use of alterative instructions for acceptable performance. adcl and sbbl are quite slow at 8 cycles for reg->reg. paddq of 32-bits within a 64-bit mmx register seems better, though the combination paddq/psrlq when propagating a carry is still a 4 cycle latency. incl and decl should be avoided, instead use add $1 and sub $1. Apparently the carry flag is not separately renamed, so incl and decl depend on all previous flags-setting instructions. shll and shrl have a 4 cycle latency, or 8 times the latency of the fastest integer instructions (addl, subl, orl, andl, and some more). shldl and shrdl seem to have 13 and 15 cycles latency, respectively. Bizarre. movq mmx -> mmx does have 6 cycle latency, as noted in the documentation. pxor/por or similar combination at 2 cycles latency can be used instead. The movq however executes in the float unit, thereby saving MMX execution resources. With the right juggling, data moves shouldn't be on a dependent chain. L1 is write-through, but the write-combining sounds like it does enough to not require explicit destination prefetching. xmm registers so far haven't found a use, but not much effort has been expended. A configure test for whether the operating system knows fxsave/fxrestor will be needed if they're used. REFERENCES Intel Pentium-4 processor manuals, http://developer.intel.com/design/pentium4/manuals "Intel Pentium 4 Processor Optimization Reference Manual", Intel, 2001, order number 248966. Available on-line: http://developer.intel.com/design/pentium4/manuals/248966.htm ---------------- Local variables: mode: text fill-column: 76 End: