|
Packit |
5c3484 |
dnl x86-64 mpn_lshiftc optimized for "Core 2".
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2007, 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 |
|
|
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 1.5
|
|
Packit |
5c3484 |
C Intel NHM 2.25 (up to about n = 260, then 1.875)
|
|
Packit |
5c3484 |
C Intel SBR 2.25
|
|
Packit |
5c3484 |
C Intel atom ?
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
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', `%rcx')
|
|
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(mpn_lshiftc)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
lea -8(rp,n,8), rp
|
|
Packit |
5c3484 |
lea -8(up,n,8), up
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(%rdx), R32(%rax)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
jne L(nb00)
|
|
Packit |
5c3484 |
L(b00): C n = 4, 8, 12, ...
|
|
Packit |
5c3484 |
mov (up), %r10
|
|
Packit |
5c3484 |
mov -8(up), %r11
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
shld R8(cnt), %r10, %rax
|
|
Packit |
5c3484 |
mov -16(up), %r8
|
|
Packit |
5c3484 |
lea 24(rp), rp
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jmp L(00)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(nb00):C n = 1, 5, 9, ...
|
|
Packit |
5c3484 |
cmp $2, R32(%rax)
|
|
Packit |
5c3484 |
jae L(nb01)
|
|
Packit |
5c3484 |
L(b01): mov (up), %r9
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
shld R8(cnt), %r9, %rax
|
|
Packit |
5c3484 |
sub $2, n
|
|
Packit |
5c3484 |
jb L(le1)
|
|
Packit |
5c3484 |
mov -8(up), %r10
|
|
Packit |
5c3484 |
mov -16(up), %r11
|
|
Packit |
5c3484 |
lea -8(up), up
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
jmp L(01)
|
|
Packit |
5c3484 |
L(le1): shl R8(cnt), %r9
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
mov %r9, (rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(nb01):C n = 2, 6, 10, ...
|
|
Packit |
5c3484 |
jne L(b11)
|
|
Packit |
5c3484 |
L(b10): mov (up), %r8
|
|
Packit |
5c3484 |
mov -8(up), %r9
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
shld R8(cnt), %r8, %rax
|
|
Packit |
5c3484 |
sub $3, n
|
|
Packit |
5c3484 |
jb L(le2)
|
|
Packit |
5c3484 |
mov -16(up), %r10
|
|
Packit |
5c3484 |
lea -16(up), up
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
jmp L(10)
|
|
Packit |
5c3484 |
L(le2): shld R8(cnt), %r9, %r8
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
shl R8(cnt), %r9
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
mov %r9, -8(rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16) C performance critical!
|
|
Packit |
5c3484 |
L(b11): C n = 3, 7, 11, ...
|
|
Packit |
5c3484 |
mov (up), %r11
|
|
Packit |
5c3484 |
mov -8(up), %r8
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
shld R8(cnt), %r11, %rax
|
|
Packit |
5c3484 |
mov -16(up), %r9
|
|
Packit |
5c3484 |
lea -24(up), up
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jb L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): shld R8(cnt), %r8, %r11
|
|
Packit |
5c3484 |
mov (up), %r10
|
|
Packit |
5c3484 |
not %r11
|
|
Packit |
5c3484 |
mov %r11, (rp)
|
|
Packit |
5c3484 |
L(10): shld R8(cnt), %r9, %r8
|
|
Packit |
5c3484 |
mov -8(up), %r11
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, -8(rp)
|
|
Packit |
5c3484 |
L(01): shld R8(cnt), %r10, %r9
|
|
Packit |
5c3484 |
mov -16(up), %r8
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
mov %r9, -16(rp)
|
|
Packit |
5c3484 |
L(00): shld R8(cnt), %r11, %r10
|
|
Packit |
5c3484 |
mov -24(up), %r9
|
|
Packit |
5c3484 |
not %r10
|
|
Packit |
5c3484 |
mov %r10, -24(rp)
|
|
Packit |
5c3484 |
add $-32, up
|
|
Packit |
5c3484 |
lea -32(rp), rp
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jnc L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): shld R8(cnt), %r8, %r11
|
|
Packit |
5c3484 |
not %r11
|
|
Packit |
5c3484 |
mov %r11, (rp)
|
|
Packit |
5c3484 |
shld R8(cnt), %r9, %r8
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, -8(rp)
|
|
Packit |
5c3484 |
shl R8(cnt), %r9
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
mov %r9, -16(rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|