|
Packit |
5c3484 |
dnl AMD64 mpn_addlshC_n -- rp[] = up[] + (vp[] << C)
|
|
Packit |
5c3484 |
dnl AMD64 mpn_rsblshC_n -- rp[] = (vp[] << C) - up[]
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2009-2012 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 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 2
|
|
Packit |
5c3484 |
C AMD K10 2
|
|
Packit |
5c3484 |
C Intel P4 ?
|
|
Packit |
5c3484 |
C Intel core2 3
|
|
Packit |
5c3484 |
C Intel NHM 2.75
|
|
Packit |
5c3484 |
C Intel SBR 2.55
|
|
Packit |
5c3484 |
C Intel atom ?
|
|
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(`vp', `%rdx')
|
|
Packit |
5c3484 |
define(`n', `%rcx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(M, eval(m4_lshift(1,LSH)))
|
|
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)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (vp), %r8
|
|
Packit |
5c3484 |
lea (,%r8,M), %r12
|
|
Packit |
5c3484 |
shr $RSH, %r8
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(n), R32(%rax)
|
|
Packit |
5c3484 |
lea (rp,n,8), rp
|
|
Packit |
5c3484 |
lea (up,n,8), up
|
|
Packit |
5c3484 |
lea (vp,n,8), vp
|
|
Packit |
5c3484 |
neg n
|
|
Packit |
5c3484 |
and $3, R8(%rax)
|
|
Packit |
5c3484 |
je L(b00)
|
|
Packit |
5c3484 |
cmp $2, R8(%rax)
|
|
Packit |
5c3484 |
jc L(b01)
|
|
Packit |
5c3484 |
je L(b10)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b11): mov 8(vp,n,8), %r10
|
|
Packit |
5c3484 |
lea (%r8,%r10,M), %r14
|
|
Packit |
5c3484 |
shr $RSH, %r10
|
|
Packit |
5c3484 |
mov 16(vp,n,8), %r11
|
|
Packit |
5c3484 |
lea (%r10,%r11,M), %r15
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
ADDSUB (up,n,8), %r12
|
|
Packit |
5c3484 |
ADCSBB 8(up,n,8), %r14
|
|
Packit |
5c3484 |
ADCSBB 16(up,n,8), %r15
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save carry for next
|
|
Packit |
5c3484 |
mov %r12, (rp,n,8)
|
|
Packit |
5c3484 |
mov %r14, 8(rp,n,8)
|
|
Packit |
5c3484 |
mov %r15, 16(rp,n,8)
|
|
Packit |
5c3484 |
add $3, n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
jmp L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b01): mov %r8, %r11
|
|
Packit |
5c3484 |
ADDSUB (up,n,8), %r12
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save carry for next
|
|
Packit |
5c3484 |
mov %r12, (rp,n,8)
|
|
Packit |
5c3484 |
add $1, n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
jmp L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b10): mov 8(vp,n,8), %r11
|
|
Packit |
5c3484 |
lea (%r8,%r11,M), %r15
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
ADDSUB (up,n,8), %r12
|
|
Packit |
5c3484 |
ADCSBB 8(up,n,8), %r15
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save carry for next
|
|
Packit |
5c3484 |
mov %r12, (rp,n,8)
|
|
Packit |
5c3484 |
mov %r15, 8(rp,n,8)
|
|
Packit |
5c3484 |
add $2, n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
jmp L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b00): mov 8(vp,n,8), %r9
|
|
Packit |
5c3484 |
mov 16(vp,n,8), %r10
|
|
Packit |
5c3484 |
jmp L(e00)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): mov 16(vp,n,8), %r10
|
|
Packit |
5c3484 |
mov (vp,n,8), %r8
|
|
Packit |
5c3484 |
mov 8(vp,n,8), %r9
|
|
Packit |
5c3484 |
lea (%r11,%r8,M), %r12
|
|
Packit |
5c3484 |
shr $RSH, %r8
|
|
Packit |
5c3484 |
L(e00): lea (%r8,%r9,M), %r13
|
|
Packit |
5c3484 |
shr $RSH, %r9
|
|
Packit |
5c3484 |
mov 24(vp,n,8), %r11
|
|
Packit |
5c3484 |
lea (%r9,%r10,M), %r14
|
|
Packit |
5c3484 |
shr $RSH, %r10
|
|
Packit |
5c3484 |
lea (%r10,%r11,M), %r15
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax) C restore carry
|
|
Packit |
5c3484 |
ADCSBB (up,n,8), %r12
|
|
Packit |
5c3484 |
ADCSBB 8(up,n,8), %r13
|
|
Packit |
5c3484 |
ADCSBB 16(up,n,8), %r14
|
|
Packit |
5c3484 |
ADCSBB 24(up,n,8), %r15
|
|
Packit |
5c3484 |
mov %r12, (rp,n,8)
|
|
Packit |
5c3484 |
mov %r13, 8(rp,n,8)
|
|
Packit |
5c3484 |
mov %r14, 16(rp,n,8)
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save carry for next
|
|
Packit |
5c3484 |
mov %r15, 24(rp,n,8)
|
|
Packit |
5c3484 |
add $4, n
|
|
Packit |
5c3484 |
js L(top)
|
|
Packit |
5c3484 |
L(end):
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifelse(ADDSUB,add,`
|
|
Packit |
5c3484 |
sub R32(%r11), R32(%rax)
|
|
Packit |
5c3484 |
neg R32(%rax)
|
|
Packit |
5c3484 |
',`
|
|
Packit |
5c3484 |
add R32(%r11), R32(%rax)
|
|
Packit |
5c3484 |
movslq R32(%rax), %rax
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
pop %r15
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|