| 1 |
dnl AMD64 mpn_addmul_1 and mpn_submul_1 optimised for Intel Haswell. |
| 2 |
|
| 3 |
dnl Contributed to the GNU project by Torbjörn Granlund. |
| 4 |
|
| 5 |
dnl Copyright 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 AMD K8,K9 n/a |
| 37 |
C AMD K10 n/a |
| 38 |
C AMD bull n/a |
| 39 |
C AMD pile n/a |
| 40 |
C AMD steam ? |
| 41 |
C AMD bobcat n/a |
| 42 |
C AMD jaguar ? |
| 43 |
C Intel P4 n/a |
| 44 |
C Intel core n/a |
| 45 |
C Intel NHM n/a |
| 46 |
C Intel SBR n/a |
| 47 |
C Intel IBR n/a |
| 48 |
C Intel HWL 2.32 |
| 49 |
C Intel BWL ? |
| 50 |
C Intel atom n/a |
| 51 |
C VIA nano n/a |
| 52 |
|
| 53 |
C The loop of this code is the result of running a code generation and |
| 54 |
C optimisation tool suite written by David Harvey and Torbjörn Granlund. |
| 55 |
|
| 56 |
C TODO |
| 57 |
C * Handle small n separately, for lower overhead. |
| 58 |
|
| 59 |
define(`rp', `%rdi') C rcx |
| 60 |
define(`up', `%rsi') C rdx |
| 61 |
define(`n_param', `%rdx') C r8 |
| 62 |
define(`v0_param',`%rcx') C r9 |
| 63 |
|
| 64 |
define(`n', `%rbp') |
| 65 |
define(`v0', `%rdx') |
| 66 |
|
| 67 |
ifdef(`OPERATION_addmul_1',` |
| 68 |
define(`ADDSUB', `add') |
| 69 |
define(`ADCSBB', `adc') |
| 70 |
define(`func', `mpn_addmul_1') |
| 71 |
') |
| 72 |
ifdef(`OPERATION_submul_1',` |
| 73 |
define(`ADDSUB', `sub') |
| 74 |
define(`ADCSBB', `sbb') |
| 75 |
define(`func', `mpn_submul_1') |
| 76 |
') |
| 77 |
|
| 78 |
ABI_SUPPORT(DOS64) |
| 79 |
ABI_SUPPORT(STD64) |
| 80 |
|
| 81 |
MULFUNC_PROLOGUE(mpn_addmul_1 mpn_submul_1) |
| 82 |
|
| 83 |
ASM_START() |
| 84 |
TEXT |
| 85 |
ALIGN(16) |
| 86 |
PROLOGUE(func) |
| 87 |
FUNC_ENTRY(4) |
| 88 |
push %rbx |
| 89 |
push %rbp |
| 90 |
push %r12 |
| 91 |
push %r13 |
| 92 |
|
| 93 |
mov n_param, n |
| 94 |
mov v0_param, v0 |
| 95 |
|
| 96 |
test $1, R8(n) |
| 97 |
jnz L(bx1) |
| 98 |
|
| 99 |
L(bx0): shr $2, n |
| 100 |
jc L(b10) |
| 101 |
|
| 102 |
L(b00): mulx( (up), %r13, %r12) |
| 103 |
mulx( 8,(up), %rbx, %rax) |
| 104 |
add %r12, %rbx |
| 105 |
adc $0, %rax |
| 106 |
mov (rp), %r12 |
| 107 |
mov 8(rp), %rcx |
| 108 |
mulx( 16,(up), %r9, %r8) |
| 109 |
lea -16(rp), rp |
| 110 |
lea 16(up), up |
| 111 |
ADDSUB %r13, %r12 |
| 112 |
jmp L(lo0) |
| 113 |
|
| 114 |
L(bx1): shr $2, n |
| 115 |
jc L(b11) |
| 116 |
|
| 117 |
L(b01): mulx( (up), %r11, %r10) |
| 118 |
jnz L(gt1) |
| 119 |
L(n1): ADDSUB %r11, (rp) |
| 120 |
mov $0, R32(%rax) |
| 121 |
adc %r10, %rax |
| 122 |
jmp L(ret) |
| 123 |
|
| 124 |
L(gt1): mulx( 8,(up), %r13, %r12) |
| 125 |
mulx( 16,(up), %rbx, %rax) |
| 126 |
lea 24(up), up |
| 127 |
add %r10, %r13 |
| 128 |
adc %r12, %rbx |
| 129 |
adc $0, %rax |
| 130 |
mov (rp), %r10 |
| 131 |
mov 8(rp), %r12 |
| 132 |
mov 16(rp), %rcx |
| 133 |
lea -8(rp), rp |
| 134 |
ADDSUB %r11, %r10 |
| 135 |
jmp L(lo1) |
| 136 |
|
| 137 |
L(b11): mulx( (up), %rbx, %rax) |
| 138 |
mov (rp), %rcx |
| 139 |
mulx( 8,(up), %r9, %r8) |
| 140 |
lea 8(up), up |
| 141 |
lea -24(rp), rp |
| 142 |
inc n C adjust n |
| 143 |
ADDSUB %rbx, %rcx |
| 144 |
jmp L(lo3) |
| 145 |
|
| 146 |
L(b10): mulx( (up), %r9, %r8) |
| 147 |
mulx( 8,(up), %r11, %r10) |
| 148 |
lea -32(rp), rp |
| 149 |
mov $0, R32(%rax) |
| 150 |
clc C clear cf |
| 151 |
jz L(end) C depends on old shift |
| 152 |
|
| 153 |
ALIGN(16) |
| 154 |
L(top): adc %rax, %r9 |
| 155 |
lea 32(rp), rp |
| 156 |
adc %r8, %r11 |
| 157 |
mulx( 16,(up), %r13, %r12) |
| 158 |
mov (rp), %r8 |
| 159 |
mulx( 24,(up), %rbx, %rax) |
| 160 |
lea 32(up), up |
| 161 |
adc %r10, %r13 |
| 162 |
adc %r12, %rbx |
| 163 |
adc $0, %rax |
| 164 |
mov 8(rp), %r10 |
| 165 |
mov 16(rp), %r12 |
| 166 |
ADDSUB %r9, %r8 |
| 167 |
mov 24(rp), %rcx |
| 168 |
mov %r8, (rp) |
| 169 |
ADCSBB %r11, %r10 |
| 170 |
L(lo1): mulx( (up), %r9, %r8) |
| 171 |
mov %r10, 8(rp) |
| 172 |
ADCSBB %r13, %r12 |
| 173 |
L(lo0): mov %r12, 16(rp) |
| 174 |
ADCSBB %rbx, %rcx |
| 175 |
L(lo3): mulx( 8,(up), %r11, %r10) |
| 176 |
mov %rcx, 24(rp) |
| 177 |
dec n |
| 178 |
jnz L(top) |
| 179 |
|
| 180 |
L(end): adc %rax, %r9 |
| 181 |
adc %r8, %r11 |
| 182 |
mov 32(rp), %r8 |
| 183 |
mov %r10, %rax |
| 184 |
adc $0, %rax |
| 185 |
mov 40(rp), %r10 |
| 186 |
ADDSUB %r9, %r8 |
| 187 |
mov %r8, 32(rp) |
| 188 |
ADCSBB %r11, %r10 |
| 189 |
mov %r10, 40(rp) |
| 190 |
adc $0, %rax |
| 191 |
|
| 192 |
L(ret): pop %r13 |
| 193 |
pop %r12 |
| 194 |
pop %rbp |
| 195 |
pop %rbx |
| 196 |
FUNC_EXIT() |
| 197 |
ret |
| 198 |
EPILOGUE() |