| 1 |
dnl AMD64 mpn_add_n, mpn_sub_n optimised for Sandy bridge, Ivy bridge, and |
| 2 |
dnl Haswell. |
| 3 |
|
| 4 |
dnl Contributed to the GNU project by Torbjörn Granlund. |
| 5 |
|
| 6 |
dnl Copyright 2003-2005, 2007, 2008, 2010-2013 Free Software Foundation, Inc. |
| 7 |
|
| 8 |
dnl This file is part of the GNU MP Library. |
| 9 |
dnl |
| 10 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify |
| 11 |
dnl it under the terms of either: |
| 12 |
dnl |
| 13 |
dnl * the GNU Lesser General Public License as published by the Free |
| 14 |
dnl Software Foundation; either version 3 of the License, or (at your |
| 15 |
dnl option) any later version. |
| 16 |
dnl |
| 17 |
dnl or |
| 18 |
dnl |
| 19 |
dnl * the GNU General Public License as published by the Free Software |
| 20 |
dnl Foundation; either version 2 of the License, or (at your option) any |
| 21 |
dnl later version. |
| 22 |
dnl |
| 23 |
dnl or both in parallel, as here. |
| 24 |
dnl |
| 25 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but |
| 26 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
| 27 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 28 |
dnl for more details. |
| 29 |
dnl |
| 30 |
dnl You should have received copies of the GNU General Public License and the |
| 31 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not, |
| 32 |
dnl see https://www.gnu.org/licenses/. |
| 33 |
|
| 34 |
include(`../config.m4') |
| 35 |
|
| 36 |
C cycles/limb |
| 37 |
C AMD K8,K9 |
| 38 |
C AMD K10 |
| 39 |
C AMD bull 1.82 average over 400-600 |
| 40 |
C AMD pile 1.83 average over 400-600 |
| 41 |
C AMD bobcat |
| 42 |
C AMD jaguar |
| 43 |
C Intel P4 |
| 44 |
C Intel core |
| 45 |
C Intel NHM |
| 46 |
C Intel SBR 1.55 fluctuates |
| 47 |
C Intel IBR 1.55 fluctuates |
| 48 |
C Intel HWL 1.33 fluctuates |
| 49 |
C Intel BWL |
| 50 |
C Intel atom |
| 51 |
C VIA nano |
| 52 |
|
| 53 |
C The loop of this code was manually written. It runs close to optimally on |
| 54 |
C Intel SBR, IBR, and HWL far as we know, except for the fluctuation problems. |
| 55 |
C It also runs slightly faster on average on AMD bull and pile. |
| 56 |
C |
| 57 |
C No micro-optimisation has been done. |
| 58 |
C |
| 59 |
C N.B.! The loop alignment padding insns are executed. If editing the code, |
| 60 |
C make sure the padding does not become excessive. It is now a 4-byte nop. |
| 61 |
|
| 62 |
define(`rp', `%rdi') C rcx |
| 63 |
define(`up', `%rsi') C rdx |
| 64 |
define(`vp', `%rdx') C r8 |
| 65 |
define(`n', `%rcx') C r9 |
| 66 |
define(`cy', `%r8') C rsp+40 (mpn_add_nc and mpn_sub_nc) |
| 67 |
|
| 68 |
ifdef(`OPERATION_add_n', ` |
| 69 |
define(ADCSBB, adc) |
| 70 |
define(func, mpn_add_n) |
| 71 |
define(func_nc, mpn_add_nc)') |
| 72 |
ifdef(`OPERATION_sub_n', ` |
| 73 |
define(ADCSBB, sbb) |
| 74 |
define(func, mpn_sub_n) |
| 75 |
define(func_nc, mpn_sub_nc)') |
| 76 |
|
| 77 |
MULFUNC_PROLOGUE(mpn_add_n mpn_add_nc mpn_sub_n mpn_sub_nc) |
| 78 |
|
| 79 |
ABI_SUPPORT(DOS64) |
| 80 |
ABI_SUPPORT(STD64) |
| 81 |
|
| 82 |
ASM_START() |
| 83 |
TEXT |
| 84 |
ALIGN(32) |
| 85 |
PROLOGUE(func) |
| 86 |
FUNC_ENTRY(4) |
| 87 |
xor %r8, %r8 |
| 88 |
|
| 89 |
L(ent): mov R32(n), R32(%rax) |
| 90 |
shr $2, n |
| 91 |
|
| 92 |
test $1, R8(%rax) |
| 93 |
jnz L(bx1) |
| 94 |
|
| 95 |
L(bx0): test $2, R8(%rax) |
| 96 |
jnz L(b10) |
| 97 |
|
| 98 |
L(b00): neg %r8 |
| 99 |
mov (up), %r8 |
| 100 |
mov 8(up), %r9 |
| 101 |
ADCSBB (vp), %r8 |
| 102 |
ADCSBB 8(vp), %r9 |
| 103 |
mov 16(up), %r10 |
| 104 |
mov 24(up), %r11 |
| 105 |
lea 32(up), up |
| 106 |
ADCSBB 16(vp), %r10 |
| 107 |
ADCSBB 24(vp), %r11 |
| 108 |
lea 32(vp), vp |
| 109 |
lea -16(rp), rp |
| 110 |
jmp L(lo0) |
| 111 |
|
| 112 |
L(b10): neg %r8 |
| 113 |
mov (up), %r10 |
| 114 |
mov 8(up), %r11 |
| 115 |
ADCSBB 0(vp), %r10 |
| 116 |
ADCSBB 8(vp), %r11 |
| 117 |
jrcxz L(e2) |
| 118 |
mov 16(up), %r8 |
| 119 |
mov 24(up), %r9 |
| 120 |
lea 16(up), up |
| 121 |
ADCSBB 16(vp), %r8 |
| 122 |
ADCSBB 24(vp), %r9 |
| 123 |
lea 16(vp), vp |
| 124 |
lea (rp), rp |
| 125 |
jmp L(lo2) |
| 126 |
|
| 127 |
L(e2): mov %r10, (rp) |
| 128 |
mov %r11, 8(rp) |
| 129 |
setc R8(%rax) |
| 130 |
FUNC_EXIT() |
| 131 |
ret |
| 132 |
|
| 133 |
L(bx1): test $2, R8(%rax) |
| 134 |
jnz L(b11) |
| 135 |
|
| 136 |
L(b01): neg %r8 |
| 137 |
mov (up), %r11 |
| 138 |
ADCSBB (vp), %r11 |
| 139 |
jrcxz L(e1) |
| 140 |
mov 8(up), %r8 |
| 141 |
mov 16(up), %r9 |
| 142 |
lea 8(up), up |
| 143 |
lea -8(rp), rp |
| 144 |
ADCSBB 8(vp), %r8 |
| 145 |
ADCSBB 16(vp), %r9 |
| 146 |
lea 8(vp), vp |
| 147 |
jmp L(lo1) |
| 148 |
|
| 149 |
L(e1): mov %r11, (rp) |
| 150 |
setc R8(%rax) |
| 151 |
FUNC_EXIT() |
| 152 |
ret |
| 153 |
|
| 154 |
L(b11): neg %r8 |
| 155 |
mov (up), %r9 |
| 156 |
ADCSBB (vp), %r9 |
| 157 |
mov 8(up), %r10 |
| 158 |
mov 16(up), %r11 |
| 159 |
lea 24(up), up |
| 160 |
ADCSBB 8(vp), %r10 |
| 161 |
ADCSBB 16(vp), %r11 |
| 162 |
lea 24(vp), vp |
| 163 |
mov %r9, (rp) |
| 164 |
lea 8(rp), rp |
| 165 |
jrcxz L(end) |
| 166 |
|
| 167 |
ALIGN(32) |
| 168 |
L(top): mov (up), %r8 |
| 169 |
mov 8(up), %r9 |
| 170 |
ADCSBB (vp), %r8 |
| 171 |
ADCSBB 8(vp), %r9 |
| 172 |
L(lo2): mov %r10, (rp) |
| 173 |
L(lo1): mov %r11, 8(rp) |
| 174 |
mov 16(up), %r10 |
| 175 |
mov 24(up), %r11 |
| 176 |
lea 32(up), up |
| 177 |
ADCSBB 16(vp), %r10 |
| 178 |
ADCSBB 24(vp), %r11 |
| 179 |
lea 32(vp), vp |
| 180 |
L(lo0): mov %r8, 16(rp) |
| 181 |
L(lo3): mov %r9, 24(rp) |
| 182 |
lea 32(rp), rp |
| 183 |
dec n |
| 184 |
jnz L(top) |
| 185 |
|
| 186 |
L(end): mov R32(n), R32(%rax) C zero rax |
| 187 |
mov %r10, (rp) |
| 188 |
mov %r11, 8(rp) |
| 189 |
setc R8(%rax) |
| 190 |
FUNC_EXIT() |
| 191 |
ret |
| 192 |
EPILOGUE() |
| 193 |
ALIGN(16) |
| 194 |
PROLOGUE(func_nc) |
| 195 |
FUNC_ENTRY(4) |
| 196 |
IFDOS(` mov 56(%rsp), %r8 ') |
| 197 |
jmp L(ent) |
| 198 |
EPILOGUE() |