|
Packit |
5c3484 |
dnl x86-64 mpn_lshift optimized for Pentium 4.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2003, 2005, 2007, 2008, 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 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 2.5
|
|
Packit |
5c3484 |
C AMD K10 ?
|
|
Packit |
5c3484 |
C Intel P4 3.29
|
|
Packit |
5c3484 |
C Intel core2 2.1 (fluctuates, presumably cache related)
|
|
Packit |
5c3484 |
C Intel corei ?
|
|
Packit |
5c3484 |
C Intel atom 14.3
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp',`%rdi')
|
|
Packit |
5c3484 |
define(`up',`%rsi')
|
|
Packit |
5c3484 |
define(`n',`%rdx')
|
|
Packit |
5c3484 |
define(`cnt',`%cl')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_lshift)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
movd R32(%rcx), %mm4
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
and $63, R32(%rcx)
|
|
Packit |
5c3484 |
movd R32(%rcx), %mm5
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea 1(n), R32(%r8)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
shr R8(%rcx), %rax C function return value
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
and $3, R32(%r8)
|
|
Packit |
5c3484 |
je L(rol) C jump for n = 3, 7, 11, ...
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dec R32(%r8)
|
|
Packit |
5c3484 |
jne L(1)
|
|
Packit |
5c3484 |
C n = 4, 8, 12, ...
|
|
Packit |
5c3484 |
movq -8(up,n,8), %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm0
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
por %mm0, %mm2
|
|
Packit |
5c3484 |
movq %mm2, -8(rp,n,8)
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jmp L(rol)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(1): dec R32(%r8)
|
|
Packit |
5c3484 |
je L(1x) C jump for n = 1, 5, 9, 13, ...
|
|
Packit |
5c3484 |
C n = 2, 6, 10, 16, ...
|
|
Packit |
5c3484 |
movq -8(up,n,8), %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm0
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
por %mm0, %mm2
|
|
Packit |
5c3484 |
movq %mm2, -8(rp,n,8)
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
L(1x):
|
|
Packit |
5c3484 |
cmp $1, n
|
|
Packit |
5c3484 |
je L(ast)
|
|
Packit |
5c3484 |
movq -8(up,n,8), %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm3
|
|
Packit |
5c3484 |
psllq %mm4, %mm3
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm0
|
|
Packit |
5c3484 |
movq -24(up,n,8), %mm1
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
por %mm0, %mm2
|
|
Packit |
5c3484 |
psrlq %mm5, %mm1
|
|
Packit |
5c3484 |
por %mm1, %mm3
|
|
Packit |
5c3484 |
movq %mm2, -8(rp,n,8)
|
|
Packit |
5c3484 |
movq %mm3, -16(rp,n,8)
|
|
Packit |
5c3484 |
sub $2, n
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(rol): movq -8(up,n,8), %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm3
|
|
Packit |
5c3484 |
psllq %mm4, %mm3
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sub $4, n C 4
|
|
Packit |
5c3484 |
jb L(end) C 2
|
|
Packit |
5c3484 |
ALIGN(32)
|
|
Packit |
5c3484 |
L(top):
|
|
Packit |
5c3484 |
C finish stuff from lsh block
|
|
Packit |
5c3484 |
movq 16(up,n,8), %mm0
|
|
Packit |
5c3484 |
movq 8(up,n,8), %mm1
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
por %mm0, %mm2
|
|
Packit |
5c3484 |
psrlq %mm5, %mm1
|
|
Packit |
5c3484 |
movq (up,n,8), %mm0
|
|
Packit |
5c3484 |
por %mm1, %mm3
|
|
Packit |
5c3484 |
movq -8(up,n,8), %mm1
|
|
Packit |
5c3484 |
movq %mm2, 24(rp,n,8)
|
|
Packit |
5c3484 |
movq %mm3, 16(rp,n,8)
|
|
Packit |
5c3484 |
C start two new rsh
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
psrlq %mm5, %mm1
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C finish stuff from rsh block
|
|
Packit |
5c3484 |
movq 8(up,n,8), %mm2
|
|
Packit |
5c3484 |
movq (up,n,8), %mm3
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
por %mm2, %mm0
|
|
Packit |
5c3484 |
psllq %mm4, %mm3
|
|
Packit |
5c3484 |
movq -8(up,n,8), %mm2
|
|
Packit |
5c3484 |
por %mm3, %mm1
|
|
Packit |
5c3484 |
movq -16(up,n,8), %mm3
|
|
Packit |
5c3484 |
movq %mm0, 8(rp,n,8)
|
|
Packit |
5c3484 |
movq %mm1, (rp,n,8)
|
|
Packit |
5c3484 |
C start two new lsh
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm3
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
jae L(top) C 2
|
|
Packit |
5c3484 |
L(end):
|
|
Packit |
5c3484 |
movq 8(up), %mm0
|
|
Packit |
5c3484 |
psrlq %mm5, %mm0
|
|
Packit |
5c3484 |
por %mm0, %mm2
|
|
Packit |
5c3484 |
movq (up), %mm1
|
|
Packit |
5c3484 |
psrlq %mm5, %mm1
|
|
Packit |
5c3484 |
por %mm1, %mm3
|
|
Packit |
5c3484 |
movq %mm2, 16(rp)
|
|
Packit |
5c3484 |
movq %mm3, 8(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(ast): movq (up), %mm2
|
|
Packit |
5c3484 |
psllq %mm4, %mm2
|
|
Packit |
5c3484 |
movq %mm2, (rp)
|
|
Packit |
5c3484 |
emms
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|