|
Packit |
5c3484 |
dnl AMD64 mpn_addlsh2_n -- rp[] = up[] + (vp[] << 2)
|
|
Packit |
5c3484 |
dnl AMD64 mpn_rsblsh2_n -- rp[] = (vp[] << 2) - up[]
|
|
Packit |
5c3484 |
dnl Optimised for Intel Atom.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjorn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 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 ?
|
|
Packit |
5c3484 |
C Intel NHM ?
|
|
Packit |
5c3484 |
C Intel SBR ?
|
|
Packit |
5c3484 |
C Intel atom 5.75
|
|
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(`LSH', 2)
|
|
Packit |
5c3484 |
define(`RSH', 62)
|
|
Packit |
5c3484 |
define(M, eval(m4_lshift(1,LSH)))
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_addlsh2_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, add)
|
|
Packit |
5c3484 |
define(ADCSBB, adc)
|
|
Packit |
5c3484 |
define(func_n, mpn_addlsh2_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_addlsh2_nc)')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsblsh2_n', `
|
|
Packit |
5c3484 |
define(ADDSUB, sub)
|
|
Packit |
5c3484 |
define(ADCSBB, sbb)
|
|
Packit |
5c3484 |
define(func_n, mpn_rsblsh2_n)
|
|
Packit |
5c3484 |
define(func_nc, mpn_rsblsh2_nc)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ABI_SUPPORT(DOS64)
|
|
Packit |
5c3484 |
ABI_SUPPORT(STD64)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
MULFUNC_PROLOGUE(mpn_addlsh2_n mpn_rsblsh2_n)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
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 R32(n), R32(%rax)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
jz L(b0) C we rely on rax = 0 at target
|
|
Packit |
5c3484 |
cmp $2, R32(%rax)
|
|
Packit |
5c3484 |
mov $0, R32(%rax)
|
|
Packit |
5c3484 |
jz L(b2)
|
|
Packit |
5c3484 |
jg L(b3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): mov (vp), %r9
|
|
Packit |
5c3484 |
lea (%rax,%r9,M), %rbp
|
|
Packit |
5c3484 |
shr $RSH, %r9
|
|
Packit |
5c3484 |
sub $1, n
|
|
Packit |
5c3484 |
lea -8(up), up
|
|
Packit |
5c3484 |
lea -8(rp), rp
|
|
Packit |
5c3484 |
jz L(cj1)
|
|
Packit |
5c3484 |
mov 8(vp), %r10
|
|
Packit |
5c3484 |
lea (%r9,%r10,M), %r9
|
|
Packit |
5c3484 |
shr $RSH, %r10
|
|
Packit |
5c3484 |
mov 16(vp), %r11
|
|
Packit |
5c3484 |
lea 24(vp), vp
|
|
Packit |
5c3484 |
mov (vp), %r8
|
|
Packit |
5c3484 |
lea (%r10,%r11,M), %r10
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
jmp L(L1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b2): lea -32(rp), rp
|
|
Packit |
5c3484 |
mov (vp), %r8
|
|
Packit |
5c3484 |
lea -32(up), up
|
|
Packit |
5c3484 |
lea (%rax,%r8,M), %rbx
|
|
Packit |
5c3484 |
shr $RSH, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
sub $2, n
|
|
Packit |
5c3484 |
jle L(end)
|
|
Packit |
5c3484 |
jmp L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): lea -24(up), up
|
|
Packit |
5c3484 |
mov (vp), %r11
|
|
Packit |
5c3484 |
lea -24(rp), rp
|
|
Packit |
5c3484 |
mov 8(vp), %r8
|
|
Packit |
5c3484 |
lea (%rax,%r11,M), %r10
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
lea 8(vp), vp
|
|
Packit |
5c3484 |
lea (%r11,%r8,M), %rbx
|
|
Packit |
5c3484 |
add $1, n
|
|
Packit |
5c3484 |
jmp L(L3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): lea -16(up), up
|
|
Packit |
5c3484 |
mov (vp), %r10
|
|
Packit |
5c3484 |
lea (%rax,%r10,M), %r9
|
|
Packit |
5c3484 |
shr $RSH, %r10
|
|
Packit |
5c3484 |
mov 8(vp), %r11
|
|
Packit |
5c3484 |
lea -16(rp), rp
|
|
Packit |
5c3484 |
mov 16(vp), %r8
|
|
Packit |
5c3484 |
lea (%r10,%r11,M), %r10
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
jmp L(L0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): lea (%r8,%r9,M), %rbp
|
|
Packit |
5c3484 |
shr $RSH, %r9
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
mov 16(vp), %r10
|
|
Packit |
5c3484 |
lea (%r9,%r10,M), %r9
|
|
Packit |
5c3484 |
shr $RSH, %r10
|
|
Packit |
5c3484 |
mov 24(vp), %r11
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
mov (vp), %r8
|
|
Packit |
5c3484 |
lea (%r10,%r11,M), %r10
|
|
Packit |
5c3484 |
shr $RSH, %r11
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
ADCSBB (up), %rbx
|
|
Packit |
5c3484 |
mov %rbx, (rp)
|
|
Packit |
5c3484 |
L(L1): ADCSBB 8(up), %rbp
|
|
Packit |
5c3484 |
mov %rbp, 8(rp)
|
|
Packit |
5c3484 |
L(L0): ADCSBB 16(up), %r9
|
|
Packit |
5c3484 |
lea (%r11,%r8,M), %rbx
|
|
Packit |
5c3484 |
mov %r9, 16(rp)
|
|
Packit |
5c3484 |
L(L3): ADCSBB 24(up), %r10
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
L(L2): shr $RSH, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
mov %r10, 24(rp)
|
|
Packit |
5c3484 |
sub $4, n
|
|
Packit |
5c3484 |
jg L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): lea (%r8,%r9,M), %rbp
|
|
Packit |
5c3484 |
shr $RSH, %r9
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
ADCSBB (up), %rbx
|
|
Packit |
5c3484 |
mov %rbx, (rp)
|
|
Packit |
5c3484 |
L(cj1): ADCSBB 8(up), %rbp
|
|
Packit |
5c3484 |
mov %rbp, 8(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`OPERATION_addlsh2_n',`
|
|
Packit |
5c3484 |
mov R32(n), R32(%rax) C zero rax
|
|
Packit |
5c3484 |
adc %r9, %rax')
|
|
Packit |
5c3484 |
ifdef(`OPERATION_rsblsh2_n',`
|
|
Packit |
5c3484 |
sbb n, %r9 C subtract 0
|
|
Packit |
5c3484 |
mov %r9, %rax')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|