|
Packit |
5c3484 |
dnl AMD64 mpn_lshift -- mpn left shift.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2003, 2005, 2007, 2009, 2011, 2012 Free Software Foundation,
|
|
Packit |
5c3484 |
dnl 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 cycles/limb cnt=1
|
|
Packit |
5c3484 |
C AMD K8,K9 2.375 1.375
|
|
Packit |
5c3484 |
C AMD K10 2.375 1.375
|
|
Packit |
5c3484 |
C Intel P4 8 10.5
|
|
Packit |
5c3484 |
C Intel core2 2.11 4.28
|
|
Packit |
5c3484 |
C Intel corei ? ?
|
|
Packit |
5c3484 |
C Intel atom 5.75 3.5
|
|
Packit |
5c3484 |
C VIA nano 3.5 2.25
|
|
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(32)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_lshift)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
cmp $1, R8(%rcx)
|
|
Packit |
5c3484 |
jne L(gen)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C For cnt=1 we want to work from lowest limb towards higher limbs.
|
|
Packit |
5c3484 |
C Check for bad overlap (up=rp is OK!) up=rp+1..rp+n-1 is bad.
|
|
Packit |
5c3484 |
C FIXME: this could surely be done more cleverly.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov rp, %rax
|
|
Packit |
5c3484 |
sub up, %rax
|
|
Packit |
5c3484 |
je L(fwd) C rp = up
|
|
Packit |
5c3484 |
shr $3, %rax
|
|
Packit |
5c3484 |
cmp n, %rax
|
|
Packit |
5c3484 |
jb L(gen)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(fwd): mov R32(n), R32(%rax)
|
|
Packit |
5c3484 |
shr $2, n
|
|
Packit |
5c3484 |
je L(e1)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(8)
|
|
Packit |
5c3484 |
nop
|
|
Packit |
5c3484 |
nop
|
|
Packit |
5c3484 |
L(t1): mov (up), %r8
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
mov 16(up), %r10
|
|
Packit |
5c3484 |
mov 24(up), %r11
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
adc %r11, %r11
|
|
Packit |
5c3484 |
mov %r11, 24(rp)
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jne L(t1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
inc R32(%rax)
|
|
Packit |
5c3484 |
dec R32(%rax)
|
|
Packit |
5c3484 |
jne L(n00)
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
L(e1): test R32(%rax), R32(%rax) C clear cy
|
|
Packit |
5c3484 |
L(n00): mov (up), %r8
|
|
Packit |
5c3484 |
dec R32(%rax)
|
|
Packit |
5c3484 |
jne L(n01)
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
L(ret): adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
L(n01): dec R32(%rax)
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
jne L(n10)
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
L(n10): mov 16(up), %r10
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
adc $-1, R32(%rax)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(gen): neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov -8(up,n,8), %rax
|
|
Packit |
5c3484 |
shr R8(%rcx), %rax C function return value
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
neg R32(%rcx) C put lsh count in cl
|
|
Packit |
5c3484 |
lea 1(n), R32(%r8)
|
|
Packit |
5c3484 |
and $3, R32(%r8)
|
|
Packit |
5c3484 |
je L(rlx) 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 |
mov -8(up,n,8), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r8
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
or %r8, %r10
|
|
Packit |
5c3484 |
mov %r10, -8(rp,n,8)
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jmp L(rll)
|
|
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 |
mov -8(up,n,8), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r8
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
or %r8, %r10
|
|
Packit |
5c3484 |
mov %r10, -8(rp,n,8)
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
neg R32(%rcx) C put lsh count in cl
|
|
Packit |
5c3484 |
L(1x):
|
|
Packit |
5c3484 |
cmp $1, n
|
|
Packit |
5c3484 |
je L(ast)
|
|
Packit |
5c3484 |
mov -8(up,n,8), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r11
|
|
Packit |
5c3484 |
shl R8(%rcx), %r11
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r8
|
|
Packit |
5c3484 |
mov -24(up,n,8), %r9
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
or %r8, %r10
|
|
Packit |
5c3484 |
shr R8(%rcx), %r9
|
|
Packit |
5c3484 |
or %r9, %r11
|
|
Packit |
5c3484 |
mov %r10, -8(rp,n,8)
|
|
Packit |
5c3484 |
mov %r11, -16(rp,n,8)
|
|
Packit |
5c3484 |
sub $2, n
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(rll): neg R32(%rcx) C put lsh count in cl
|
|
Packit |
5c3484 |
L(rlx): mov -8(up,n,8), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r11
|
|
Packit |
5c3484 |
shl R8(%rcx), %r11
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sub $4, n C 4
|
|
Packit |
5c3484 |
jb L(end) C 2
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top):
|
|
Packit |
5c3484 |
C finish stuff from lsh block
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov 16(up,n,8), %r8
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r9
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
or %r8, %r10
|
|
Packit |
5c3484 |
shr R8(%rcx), %r9
|
|
Packit |
5c3484 |
or %r9, %r11
|
|
Packit |
5c3484 |
mov %r10, 24(rp,n,8)
|
|
Packit |
5c3484 |
mov %r11, 16(rp,n,8)
|
|
Packit |
5c3484 |
C start two new rsh
|
|
Packit |
5c3484 |
mov 0(up,n,8), %r8
|
|
Packit |
5c3484 |
mov -8(up,n,8), %r9
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
shr R8(%rcx), %r9
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C finish stuff from rsh block
|
|
Packit |
5c3484 |
neg R32(%rcx) C put lsh count in cl
|
|
Packit |
5c3484 |
mov 8(up,n,8), %r10
|
|
Packit |
5c3484 |
mov 0(up,n,8), %r11
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
or %r10, %r8
|
|
Packit |
5c3484 |
shl R8(%rcx), %r11
|
|
Packit |
5c3484 |
or %r11, %r9
|
|
Packit |
5c3484 |
mov %r8, 8(rp,n,8)
|
|
Packit |
5c3484 |
mov %r9, 0(rp,n,8)
|
|
Packit |
5c3484 |
C start two new lsh
|
|
Packit |
5c3484 |
mov -8(up,n,8), %r10
|
|
Packit |
5c3484 |
mov -16(up,n,8), %r11
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r11
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jae L(top) C 2
|
|
Packit |
5c3484 |
L(end):
|
|
Packit |
5c3484 |
neg R32(%rcx) C put rsh count in cl
|
|
Packit |
5c3484 |
mov 8(up), %r8
|
|
Packit |
5c3484 |
shr R8(%rcx), %r8
|
|
Packit |
5c3484 |
or %r8, %r10
|
|
Packit |
5c3484 |
mov (up), %r9
|
|
Packit |
5c3484 |
shr R8(%rcx), %r9
|
|
Packit |
5c3484 |
or %r9, %r11
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
mov %r11, 8(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
neg R32(%rcx) C put lsh count in cl
|
|
Packit |
5c3484 |
L(ast): mov (up), %r10
|
|
Packit |
5c3484 |
shl R8(%rcx), %r10
|
|
Packit |
5c3484 |
mov %r10, (rp)
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|