|
Packit |
5c3484 |
dnl AMD64 mpn_addmul_1 and mpn_submul_1 optimised for AMD Bulldozer.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2003-2005, 2007, 2008, 2011, 2012 Free Software Foundation, Inc.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl This file is part of the GNU MP Library.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify
|
|
Packit |
5c3484 |
dnl it under the terms of either:
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU Lesser General Public License as published by the Free
|
|
Packit |
5c3484 |
dnl Software Foundation; either version 3 of the License, or (at your
|
|
Packit |
5c3484 |
dnl option) any later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl * the GNU General Public License as published by the Free Software
|
|
Packit |
5c3484 |
dnl Foundation; either version 2 of the License, or (at your option) any
|
|
Packit |
5c3484 |
dnl later version.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl or both in parallel, as here.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but
|
|
Packit |
5c3484 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
|
|
Packit |
5c3484 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
Packit |
5c3484 |
dnl for more details.
|
|
Packit |
5c3484 |
dnl
|
|
Packit |
5c3484 |
dnl You should have received copies of the GNU General Public License and the
|
|
Packit |
5c3484 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not,
|
|
Packit |
5c3484 |
dnl see https://www.gnu.org/licenses/.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
include(`../config.m4')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9
|
|
Packit |
5c3484 |
C AMD K10
|
|
Packit |
5c3484 |
C AMD bd1 4.5-4.7
|
|
Packit |
5c3484 |
C AMD bobcat
|
|
Packit |
5c3484 |
C Intel P4
|
|
Packit |
5c3484 |
C Intel core2
|
|
Packit |
5c3484 |
C Intel NHM
|
|
Packit |
5c3484 |
C Intel SBR
|
|
Packit |
5c3484 |
C Intel atom
|
|
Packit |
5c3484 |
C VIA nano
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C The loop of this code is the result of running a code generation and
|
|
Packit |
5c3484 |
C optimisation tool suite written by David Harvey and Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Try to make loop run closer to 4 c/l.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`rp', `%rdi') C rcx
|
|
Packit |
5c3484 |
define(`up', `%rsi') C rdx
|
|
Packit |
5c3484 |
define(`n_param', `%rdx') C r8
|
|
Packit |
5c3484 |
define(`v0', `%rcx') C r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`n', `%r11')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_addmul_1',`
|
|
Packit |
5c3484 |
define(`ADDSUB', `add')
|
|
Packit |
5c3484 |
define(`func', `mpn_addmul_1')
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_submul_1',`
|
|
Packit |
5c3484 |
define(`ADDSUB', `sub')
|
|
Packit |
5c3484 |
define(`func', `mpn_submul_1')
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_addmul_1 mpn_submul_1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
IFDOS(` define(`up', ``%rsi'') ') dnl
|
|
Packit |
5c3484 |
IFDOS(` define(`rp', ``%rcx'') ') dnl
|
|
Packit |
5c3484 |
IFDOS(` define(`v0', ``%r9'') ') dnl
|
|
Packit |
5c3484 |
IFDOS(` define(`r9', ``rdi'') ') dnl
|
|
Packit |
5c3484 |
IFDOS(` define(`n', ``%r8'') ') dnl
|
|
Packit |
5c3484 |
IFDOS(` define(`r8', ``r11'') ') dnl
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(func)
|
|
Packit |
5c3484 |
IFDOS(``push %rsi '')
|
|
Packit |
5c3484 |
IFDOS(``push %rdi '')
|
|
Packit |
5c3484 |
IFDOS(``mov %rdx, %rsi '')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax C read first u limb early
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
IFSTD(` mov n_param, %rbx ') C move away n from rdx, mul uses it
|
|
Packit |
5c3484 |
IFDOS(` mov n, %rbx ')
|
|
Packit |
5c3484 |
mul v0
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
IFSTD(` mov %rbx, n ')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
and $3, R32(%rbx)
|
|
Packit |
5c3484 |
lea -16(rp,n,8), rp
|
|
Packit |
5c3484 |
jz L(b0)
|
|
Packit |
5c3484 |
cmp $2, R32(%rbx)
|
|
Packit |
5c3484 |
jb L(b1)
|
|
Packit |
5c3484 |
jz L(b2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): mov $0, R32(%r8)
|
|
Packit |
5c3484 |
mov %rax, %rbx
|
|
Packit |
5c3484 |
mov $0, R32(%r9)
|
|
Packit |
5c3484 |
mov 8(up), %rax
|
|
Packit |
5c3484 |
mov %rdx, %r10
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
not n
|
|
Packit |
5c3484 |
jmp L(L3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): mov $0, R32(%r10)
|
|
Packit |
5c3484 |
mov %rax, %r8
|
|
Packit |
5c3484 |
mov %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(up), %rax
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
jmp L(L0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): cmp $1, n
|
|
Packit |
5c3484 |
jz L(n1)
|
|
Packit |
5c3484 |
mov %rax, %r9
|
|
Packit |
5c3484 |
mov 8(up), %rax
|
|
Packit |
5c3484 |
mov %rdx, %r8
|
|
Packit |
5c3484 |
mov $0, R32(%rbx)
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
inc n
|
|
Packit |
5c3484 |
jmp L(L1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b2): mov $0, R32(%rbx)
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mov %rdx, %r9
|
|
Packit |
5c3484 |
mov 8(up), %rax
|
|
Packit |
5c3484 |
mov $0, R32(%r8)
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
jns L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(top): mul v0
|
|
Packit |
5c3484 |
ADDSUB %r10, (rp,n,8)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov (up,n,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r8
|
|
Packit |
5c3484 |
L(L1): mul v0
|
|
Packit |
5c3484 |
mov $0, R32(%r10)
|
|
Packit |
5c3484 |
ADDSUB %r9, 8(rp,n,8)
|
|
Packit |
5c3484 |
adc %rax, %r8
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(up,n,8), %rax
|
|
Packit |
5c3484 |
L(L0): mul v0
|
|
Packit |
5c3484 |
ADDSUB %r8, 16(rp,n,8)
|
|
Packit |
5c3484 |
mov $0, R32(%r8)
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
mov $0, R32(%r9)
|
|
Packit |
5c3484 |
mov 16(up,n,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
L(L3): mul v0
|
|
Packit |
5c3484 |
ADDSUB %rbx, 24(rp,n,8)
|
|
Packit |
5c3484 |
mov $0, R32(%rbx)
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mov 24(up,n,8), %rax
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): mul v0
|
|
Packit |
5c3484 |
ADDSUB %r10, (rp)
|
|
Packit |
5c3484 |
adc %r9, %rax
|
|
Packit |
5c3484 |
adc %r8, %rdx
|
|
Packit |
5c3484 |
L(n1): ADDSUB %rax, 8(rp)
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, %rax
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
IFDOS(``pop %rdi '')
|
|
Packit |
5c3484 |
IFDOS(``pop %rsi '')
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
ASM_END()
|
|
rpm-build |
01f633 |
CF_PROT
|