arch/powerpc/lib/copypage_power7.S

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License as published by
   4  * the Free Software Foundation; either version 2 of the License, or
   5  * (at your option) any later version.
   6  *
   7  * This program is distributed in the hope that it will be useful,
   8  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   9  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  10  * GNU General Public License for more details.
  11  *
  12  * You should have received a copy of the GNU General Public License
  13  * along with this program; if not, write to the Free Software
  14  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  15  *
  16  * Copyright (C) IBM Corporation, 2012
  17  *
  18  * Author: Anton Blanchard <anton@au.ibm.com>
  19  */
  20 #include <asm/page.h>
  21 #include <asm/ppc_asm.h>
  22
  23 _GLOBAL(copypage_power7)
  24         /*
  25          * We prefetch both the source and destination using enhanced touch
  26          * instructions. We use a stream ID of 0 for the load side and
  27          * 1 for the store side. Since source and destination are page
  28          * aligned we don't need to clear the bottom 7 bits of either
  29          * address.
  30          */
  31         ori     r9,r3,1         /* stream=1 => to */
  32
  33 #ifdef CONFIG_PPC_64K_PAGES
  34         lis     r7,0x0E01       /* depth=7
  35                                  * units/cachelines=512 */
  36 #else
  37         lis     r7,0x0E00       /* depth=7 */
  38         ori     r7,r7,0x1000    /* units/cachelines=32 */
  39 #endif
  40         ori     r10,r7,1        /* stream=1 */
  41
  42         lis     r8,0x8000       /* GO=1 */
  43         clrldi  r8,r8,32
  44
  45 .machine push
  46 .machine "power4"
  47         /* setup read stream 0  */
  48         dcbt    0,r4,0b01000    /* addr from */
  49         dcbt    0,r7,0b01010   /* length and depth from */
  50         /* setup write stream 1 */
  51         dcbtst  0,r9,0b01000   /* addr to */
  52         dcbtst  0,r10,0b01010  /* length and depth to */
  53         eieio
  54         dcbt    0,r8,0b01010    /* all streams GO */
  55 .machine pop
  56
  57 #ifdef CONFIG_ALTIVEC
  58         mflr    r0
  59         std     r3,-STACKFRAMESIZE+STK_REG(R31)(r1)
  60         std     r4,-STACKFRAMESIZE+STK_REG(R30)(r1)
  61         std     r0,16(r1)
  62         stdu    r1,-STACKFRAMESIZE(r1)
  63         bl      enter_vmx_copy
  64         cmpwi   r3,0
  65         ld      r0,STACKFRAMESIZE+16(r1)
  66         ld      r3,STK_REG(R31)(r1)
  67         ld      r4,STK_REG(R30)(r1)
  68         mtlr    r0
  69
  70         li      r0,(PAGE_SIZE/128)
  71         mtctr   r0
  72
  73         beq     .Lnonvmx_copy
  74
  75         addi    r1,r1,STACKFRAMESIZE
  76
  77         li      r6,16
  78         li      r7,32
  79         li      r8,48
  80         li      r9,64
  81         li      r10,80
  82         li      r11,96
  83         li      r12,112
  84
  85         .align  5
  86 1:      lvx     v7,0,r4
  87         lvx     v6,r4,r6
  88         lvx     v5,r4,r7
  89         lvx     v4,r4,r8
  90         lvx     v3,r4,r9
  91         lvx     v2,r4,r10
  92         lvx     v1,r4,r11
  93         lvx     v0,r4,r12
  94         addi    r4,r4,128
  95         stvx    v7,0,r3
  96         stvx    v6,r3,r6
  97         stvx    v5,r3,r7
  98         stvx    v4,r3,r8
  99         stvx    v3,r3,r9
 100         stvx    v2,r3,r10
 101         stvx    v1,r3,r11
 102         stvx    v0,r3,r12
 103         addi    r3,r3,128
 104         bdnz    1b
 105
 106         b       exit_vmx_copy           /* tail call optimise */
 107
 108 #else
 109         li      r0,(PAGE_SIZE/128)
 110         mtctr   r0
 111
 112         stdu    r1,-STACKFRAMESIZE(r1)
 113 #endif
 114
 115 .Lnonvmx_copy:
 116         std     r14,STK_REG(R14)(r1)
 117         std     r15,STK_REG(R15)(r1)
 118         std     r16,STK_REG(R16)(r1)
 119         std     r17,STK_REG(R17)(r1)
 120         std     r18,STK_REG(R18)(r1)
 121         std     r19,STK_REG(R19)(r1)
 122         std     r20,STK_REG(R20)(r1)
 123
 124 1:      ld      r0,0(r4)
 125         ld      r5,8(r4)
 126         ld      r6,16(r4)
 127         ld      r7,24(r4)
 128         ld      r8,32(r4)
 129         ld      r9,40(r4)
 130         ld      r10,48(r4)
 131         ld      r11,56(r4)
 132         ld      r12,64(r4)
 133         ld      r14,72(r4)
 134         ld      r15,80(r4)
 135         ld      r16,88(r4)
 136         ld      r17,96(r4)
 137         ld      r18,104(r4)
 138         ld      r19,112(r4)
 139         ld      r20,120(r4)
 140         addi    r4,r4,128
 141         std     r0,0(r3)
 142         std     r5,8(r3)
 143         std     r6,16(r3)
 144         std     r7,24(r3)
 145         std     r8,32(r3)
 146         std     r9,40(r3)
 147         std     r10,48(r3)
 148         std     r11,56(r3)
 149         std     r12,64(r3)
 150         std     r14,72(r3)
 151         std     r15,80(r3)
 152         std     r16,88(r3)
 153         std     r17,96(r3)
 154         std     r18,104(r3)
 155         std     r19,112(r3)
 156         std     r20,120(r3)
 157         addi    r3,r3,128
 158         bdnz    1b
 159
 160         ld      r14,STK_REG(R14)(r1)
 161         ld      r15,STK_REG(R15)(r1)
 162         ld      r16,STK_REG(R16)(r1)
 163         ld      r17,STK_REG(R17)(r1)
 164         ld      r18,STK_REG(R18)(r1)
 165         ld      r19,STK_REG(R19)(r1)
 166         ld      r20,STK_REG(R20)(r1)
 167         addi    r1,r1,STACKFRAMESIZE
 168         blr