1Copyright 2011 Free Software Foundation, Inc. 2 3This file is part of the GNU MP Library. 4 5The GNU MP Library is free software; you can redistribute it and/or modify 6it under the terms of either: 7 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. 11 12or 13 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 16 later version. 17 18or both in parallel, as here. 19 20The GNU MP Library is distributed in the hope that it will be useful, but 21WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 22or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 23for more details. 24 25You should have received copies of the GNU General Public License and the 26GNU Lesser General Public License along with the GNU MP Library. If not, 27see https://www.gnu.org/licenses/. 28 29 30 31There are 5 generations of 64-but s390 processors, z900, z990, z9, 32z10, and z196. The current GMP code was optimised for the two oldest, 33z900 and z990. 34 35 36mpn_copyi 37 38This code makes use of a loop around MVC. It almost surely runs very 39close to optimally. A small improvement could be done by using one 40MVC for size 256 bytes, now we use two (we use an extra MVC when 41copying any multiple of 256 bytes). 42 43 44mpn_copyd 45 46We have tried several feed-in variants here, branch tree, jump table 47and computed goto. The fastest (on z990) turned out to be computed 48goto. 49 50An approach not tried is EX of LMG and STMG, modifying the register set 51on-the-fly. Using that trick, we could completely avoid using 52separate feed-in paths. 53 54 55mpn_lshift, mpn_rshift 56 57The current code runs at pipeline decode bandwidth on z990. 58 59 60mpn_add_n, mpn_sub_n 61 62The current code is 4-way unrolled. It should be unrolled more, at 63least 8x, in order to reach 2.5 c/l. 64 65 66mpn_mul_1, mpn_addmul_1, mpn_submul_1 67 68The current code is very naive, but due to the non-pipelined nature of 69MLGR on z900 and z990, more sophisticated code would not gain much. 70 71On z10 one would need to cluster at least 4 MLGR together, in order to 72reduce stalling. 73 74On z196, one surely want to use unrolling and pipelining, to perhaps 75reach around 12 c/l. A major issue here and on z10 is ALCGR's 3 cycle 76stalling. 77 78 79mpn_mul_2, mpn_addmul_2 80 81At least for older machines (z900, z990) with very slow MLGR, we 82should use Karatsuba's algorithm on 2-limb units, making mul_2 and 83addmul_2 the main multiplication primitives. The newer machines might 84benefit less from this approach, perhaps in particular z10, where MLGR 85clustering is more important. 86 87With Karatsuba, one could hope for around 16 cycles per accumulated 88128 cross product, on z990. 89