source/libs/gmp/gmp-src/mpn/powerpc64/sec_tabselect.asm

   1 dnl  PowerPC-64 mpn_sec_tabselect.
   2
   3 dnl  Contributed to the GNU project by Torbjörn Granlund.
   4
   5 dnl  Copyright 2011-2013 Free Software Foundation, Inc.
   6
   7 dnl  This file is part of the GNU MP Library.
   8 dnl
   9 dnl  The GNU MP Library is free software; you can redistribute it and/or modify
  10 dnl  it under the terms of either:
  11 dnl
  12 dnl    * the GNU Lesser General Public License as published by the Free
  13 dnl      Software Foundation; either version 3 of the License, or (at your
  14 dnl      option) any later version.
  15 dnl
  16 dnl  or
  17 dnl
  18 dnl    * the GNU General Public License as published by the Free Software
  19 dnl      Foundation; either version 2 of the License, or (at your option) any
  20 dnl      later version.
  21 dnl
  22 dnl  or both in parallel, as here.
  23 dnl
  24 dnl  The GNU MP Library is distributed in the hope that it will be useful, but
  25 dnl  WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  26 dnl  or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  27 dnl  for more details.
  28 dnl
  29 dnl  You should have received copies of the GNU General Public License and the
  30 dnl  GNU Lesser General Public License along with the GNU MP Library.  If not,
  31 dnl  see https://www.gnu.org/licenses/.
  32
  33 include(`../config.m4')
  34
  35 C                   cycles/limb
  36 C POWER3/PPC630          1.75
  37 C POWER4/PPC970          2.0
  38 C POWER5                 ?
  39 C POWER6                 5.0
  40 C POWER7                 1.75
  41
  42 define(`rp',     `r3')
  43 define(`tp',     `r4')
  44 define(`n',      `r5')
  45 define(`nents',  `r6')
  46 define(`which',  `r7')
  47
  48 define(`i',      `r8')
  49 define(`j',      `r9')
  50 define(`stride', `r12')
  51 define(`mask',   `r11')
  52
  53
  54 ASM_START()
  55 PROLOGUE(mpn_sec_tabselect)
  56         addic.  j, n, -4                C outer loop induction variable
  57         std     r31, -8(r1)
  58         std     r30, -16(r1)
  59         std     r29, -24(r1)
  60         std     r28, -32(r1)
  61         std     r27, -40(r1)
  62         sldi    stride, n, 3
  63
  64         blt     cr0, L(outer_end)
  65 L(outer_top):
  66         mtctr   nents
  67         mr      r10, tp
  68         li      r28, 0
  69         li      r29, 0
  70         li      r30, 0
  71         li      r31, 0
  72         addic.  j, j, -4                C outer loop induction variable
  73         mr      i, which
  74
  75         ALIGN(16)
  76 L(top): addic   i, i, -1                C set carry iff i != 0
  77         subfe   mask, mask, mask
  78         ld      r0, 0(tp)
  79         ld      r27, 8(tp)
  80         and     r0, r0, mask
  81         and     r27, r27, mask
  82         or      r28, r28, r0
  83         or      r29, r29, r27
  84         ld      r0, 16(tp)
  85         ld      r27, 24(tp)
  86         and     r0, r0, mask
  87         and     r27, r27, mask
  88         or      r30, r30, r0
  89         or      r31, r31, r27
  90         add     tp, tp, stride
  91         bdnz    L(top)
  92
  93         std     r28, 0(rp)
  94         std     r29, 8(rp)
  95         std     r30, 16(rp)
  96         std     r31, 24(rp)
  97         addi    tp, r10, 32
  98         addi    rp, rp, 32
  99         bge     cr0, L(outer_top)
 100 L(outer_end):
 101
 102         rldicl. r0, n, 63, 63
 103         beq     cr0, L(b0x)
 104 L(b1x): mtctr   nents
 105         mr      r10, tp
 106         li      r28, 0
 107         li      r29, 0
 108         mr      i, which
 109         ALIGN(16)
 110 L(tp2): addic   i, i, -1
 111         subfe   mask, mask, mask
 112         ld      r0, 0(tp)
 113         ld      r27, 8(tp)
 114         and     r0, r0, mask
 115         and     r27, r27, mask
 116         or      r28, r28, r0
 117         or      r29, r29, r27
 118         add     tp, tp, stride
 119         bdnz    L(tp2)
 120         std     r28, 0(rp)
 121         std     r29, 8(rp)
 122         addi    tp, r10, 16
 123         addi    rp, rp, 16
 124
 125 L(b0x): rldicl. r0, n, 0, 63
 126         beq     cr0, L(b00)
 127 L(b01): mtctr   nents
 128         mr      r10, tp
 129         li      r28, 0
 130         mr      i, which
 131         ALIGN(16)
 132 L(tp1): addic   i, i, -1
 133         subfe   mask, mask, mask
 134         ld      r0, 0(tp)
 135         and     r0, r0, mask
 136         or      r28, r28, r0
 137         add     tp, tp, stride
 138         bdnz    L(tp1)
 139         std     r28, 0(rp)
 140
 141 L(b00): ld      r31, -8(r1)
 142         ld      r30, -16(r1)
 143         ld      r29, -24(r1)
 144         ld      r28, -32(r1)
 145         ld      r27, -40(r1)
 146         blr
 147 EPILOGUE()