|
Packit |
5c3484 |
dnl X86-64 mpn_rsh1add_n, mpn_rsh1sub_n optimised for Intel Conroe/Penryn.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2003, 2005, 2009, 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 Intel P4 ?
|
|
Packit |
5c3484 |
C Intel core2 3.05
|
|
Packit |
5c3484 |
C Intel NHM 3.3
|
|
Packit |
5c3484 |
C Intel SBR 2.5
|
|
Packit |
5c3484 |
C Intel atom ?
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Loopmix to approach 2.5 c/l on NHM.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp', `%rdi')
|
|
Packit |
5c3484 |
define(`up', `%rsi')
|
|
Packit |
5c3484 |
define(`vp', `%rdx')
|
|
Packit |
5c3484 |
define(`n', `%rcx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsh1add_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, add)
|
|
Packit |
5c3484 |
define(ADCSBB, adc)
|
|
Packit |
5c3484 |
define(func_n, mpn_rsh1add_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_rsh1add_nc)')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsh1sub_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, sub)
|
|
Packit |
5c3484 |
define(ADCSBB, sbb)
|
|
Packit |
5c3484 |
define(func_n, mpn_rsh1sub_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_rsh1sub_nc)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_rsh1add_n mpn_rsh1add_nc mpn_rsh1sub_n mpn_rsh1sub_nc)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(func_nc)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
IFDOS(` mov 56(%rsp), %r8 ')
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
neg %r8 C set C flag from parameter
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
ADCSBB (vp), %r8
|
|
Packit |
5c3484 |
jmp L(ent)
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
PROLOGUE(func_n)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
ADDSUB (vp), %r8
|
|
Packit |
5c3484 |
L(ent): sbb R32(%rbx), R32(%rbx) C save cy
|
|
Packit |
5c3484 |
mov %r8, %rax
|
|
Packit |
5c3484 |
and $1, R32(%rax) C return value
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
lea (vp,n,8), vp
|
|
Packit |
5c3484 |
lea (rp,n,8), rp
|
|
Packit |
5c3484 |
mov R32(n), R32(%rbp)
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
and $3, R32(%rbp)
|
|
Packit |
5c3484 |
jz L(b0)
|
|
Packit |
5c3484 |
cmp $2, R32(%rbp)
|
|
Packit |
5c3484 |
jae L(n1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): mov %r8, %rbp
|
|
Packit |
5c3484 |
inc n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
jmp L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(n1): jnz L(b3)
|
|
Packit |
5c3484 |
add R32(%rbx), R32(%rbx) C restore cy
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r11
|
|
Packit |
5c3484 |
ADCSBB 8(vp,n,8), %r11
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx) C save cy
|
|
Packit |
5c3484 |
mov %r8, %r10
|
|
Packit |
5c3484 |
add $-2, n
|
|
Packit |
5c3484 |
jmp L(2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): add R32(%rbx), R32(%rbx) C restore cy
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r10
|
|
Packit |
5c3484 |
mov 16(up,n,8), %r11
|
|
Packit |
5c3484 |
ADCSBB 8(vp,n,8), %r10
|
|
Packit |
5c3484 |
ADCSBB 16(vp,n,8), %r11
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx) C save cy
|
|
Packit |
5c3484 |
mov %r8, %r9
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jmp L(3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): add R32(%rbx), R32(%rbx) C restore cy
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r9
|
|
Packit |
5c3484 |
mov 16(up,n,8), %r10
|
|
Packit |
5c3484 |
mov 24(up,n,8), %r11
|
|
Packit |
5c3484 |
ADCSBB 8(vp,n,8), %r9
|
|
Packit |
5c3484 |
ADCSBB 16(vp,n,8), %r10
|
|
Packit |
5c3484 |
ADCSBB 24(vp,n,8), %r11
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx) C save cy
|
|
Packit |
5c3484 |
jmp L(4)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(top): add R32(%rbx), R32(%rbx) C restore cy
|
|
Packit |
5c3484 |
mov (up,n,8), %r8
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r9
|
|
Packit |
5c3484 |
mov 16(up,n,8), %r10
|
|
Packit |
5c3484 |
mov 24(up,n,8), %r11
|
|
Packit |
5c3484 |
ADCSBB (vp,n,8), %r8
|
|
Packit |
5c3484 |
ADCSBB 8(vp,n,8), %r9
|
|
Packit |
5c3484 |
ADCSBB 16(vp,n,8), %r10
|
|
Packit |
5c3484 |
ADCSBB 24(vp,n,8), %r11
|
|
Packit |
5c3484 |
sbb R32(%rbx), R32(%rbx) C save cy
|
|
Packit |
5c3484 |
shrd $1, %r8, %rbp
|
|
Packit |
5c3484 |
mov %rbp, -8(rp,n,8)
|
|
Packit |
5c3484 |
L(4): shrd $1, %r9, %r8
|
|
Packit |
5c3484 |
mov %r8, (rp,n,8)
|
|
Packit |
5c3484 |
L(3): shrd $1, %r10, %r9
|
|
Packit |
5c3484 |
mov %r9, 8(rp,n,8)
|
|
Packit |
5c3484 |
L(2): shrd $1, %r11, %r10
|
|
Packit |
5c3484 |
mov %r10, 16(rp,n,8)
|
|
Packit |
5c3484 |
L(1): add $4, n
|
|
Packit |
5c3484 |
mov %r11, %rbp
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): shrd $1, %rbx, %rbp
|
|
Packit |
5c3484 |
mov %rbp, -8(rp)
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|