1dnl HP-PA 32-bit mpn_sqr_diagonal optimized for the PA8x00. 2 3dnl Copyright 2001, 2002 Free Software Foundation, Inc. 4 5dnl This file is part of the GNU MP Library. 6 7dnl The GNU MP Library is free software; you can redistribute it and/or modify 8dnl it under the terms of the GNU Lesser General Public License as published 9dnl by the Free Software Foundation; either version 3 of the License, or (at 10dnl your option) any later version. 11 12dnl The GNU MP Library is distributed in the hope that it will be useful, but 13dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY 14dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public 15dnl License for more details. 16 17dnl You should have received a copy of the GNU Lesser General Public License 18dnl along with the GNU MP Library. If not, see http://www.gnu.org/licenses/. 19 20include(`../config.m4') 21 22C This code runs at 6 cycles/limb on the PA7100 and 2 cycles/limb on PA8x00. 23C The 2-way unrolling is actually not helping the PA7100. 24 25C INPUT PARAMETERS 26define(`rp',`%r26') 27define(`up',`%r25') 28define(`n',`%r24') 29 30ASM_START() 31PROLOGUE(mpn_sqr_diagonal) 32 33 fldws,ma 4(up),%fr4r 34 addib,= -1,n,L(end1) 35 ldo 4(rp),rp 36 37 fldws,ma 4(up),%fr6r 38 addib,= -1,n,L(end2) 39 xmpyu %fr4r,%fr4r,%fr5 40 41 fldws,ma 4(up),%fr4r 42 addib,= -1,n,L(end3) 43 xmpyu %fr6r,%fr6r,%fr7 44 45 46LDEF(loop) 47 fldws,ma 4(up),%fr6r 48 fstws %fr5r,-4(rp) 49 fstws,ma %fr5l,8(rp) 50 addib,= -1,n,L(exite) 51 xmpyu %fr4r,%fr4r,%fr5 52 fldws,ma 4(up),%fr4r 53 fstws %fr7r,-4(rp) 54 fstws,ma %fr7l,8(rp) 55 addib,<> -1,n,L(loop) 56 xmpyu %fr6r,%fr6r,%fr7 57 58LDEF(exito) 59 fstws %fr5r,-4(rp) 60 fstws %fr5l,0(rp) 61 xmpyu %fr4r,%fr4r,%fr5 62 fstws %fr7r,4(rp) 63 fstws %fr7l,8(rp) 64 fstws,mb %fr5r,12(rp) 65 bv 0(%r2) 66 fstws %fr5l,4(rp) 67 68LDEF(exite) 69 fstws %fr7r,-4(rp) 70 fstws %fr7l,0(rp) 71 xmpyu %fr6r,%fr6r,%fr7 72 fstws %fr5r,4(rp) 73 fstws %fr5l,8(rp) 74 fstws,mb %fr7r,12(rp) 75 bv 0(%r2) 76 fstws %fr7l,4(rp) 77 78LDEF(end1) 79 xmpyu %fr4r,%fr4r,%fr5 80 fstws %fr5r,-4(rp) 81 bv 0(%r2) 82 fstws,ma %fr5l,8(rp) 83 84LDEF(end2) 85 xmpyu %fr6r,%fr6r,%fr7 86 fstws %fr5r,-4(rp) 87 fstws %fr5l,0(rp) 88 fstws %fr7r,4(rp) 89 bv 0(%r2) 90 fstws %fr7l,8(rp) 91 92LDEF(end3) 93 fstws %fr5r,-4(rp) 94 fstws %fr5l,0(rp) 95 xmpyu %fr4r,%fr4r,%fr5 96 fstws %fr7r,4(rp) 97 fstws %fr7l,8(rp) 98 fstws,mb %fr5r,12(rp) 99 bv 0(%r2) 100 fstws %fr5l,4(rp) 101EPILOGUE(mpn_sqr_diagonal) 102