|
Packit |
5c3484 |
dnl X86-64 mpn_redc_1 optimised for AMD K8-K10.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjörn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2004, 2008, 2013 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 AMD bull ?
|
|
Packit |
5c3484 |
C AMD pile ?
|
|
Packit |
5c3484 |
C AMD steam ?
|
|
Packit |
5c3484 |
C AMD bobcat ?
|
|
Packit |
5c3484 |
C AMD jaguar ?
|
|
Packit |
5c3484 |
C Intel P4 ?
|
|
Packit |
5c3484 |
C Intel core ?
|
|
Packit |
5c3484 |
C Intel NHM ?
|
|
Packit |
5c3484 |
C Intel SBR ?
|
|
Packit |
5c3484 |
C Intel IBR ?
|
|
Packit |
5c3484 |
C Intel HWL ?
|
|
Packit |
5c3484 |
C Intel BWL ?
|
|
Packit |
5c3484 |
C Intel atom ?
|
|
Packit |
5c3484 |
C VIA nano ?
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C The inner loops of this code are the result of running a code generation and
|
|
Packit |
5c3484 |
C optimisation tool suite written by David Harvey and Torbjörn Granlund.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C TODO
|
|
Packit |
5c3484 |
C * Micro-optimise, none performed thus far.
|
|
Packit |
5c3484 |
C * This looks different from other current redc_1.asm variants. Consider
|
|
Packit |
5c3484 |
C adapting this to the mainstream style.
|
|
Packit |
5c3484 |
C * Is this code really faster than more approaches which compute q0 later?
|
|
Packit |
5c3484 |
C Is the use of a jump jump table faster? Or is the edge of this due to the
|
|
Packit |
5c3484 |
C inlined add_n code?
|
|
Packit |
5c3484 |
C * Put initial m[0] x q0 computation in header.
|
|
Packit |
5c3484 |
C * Put basecases at the file's end, single them out before the pushes.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`rp', `%rdi') C rcx
|
|
Packit |
5c3484 |
define(`up', `%rsi') C rdx
|
|
Packit |
5c3484 |
define(`mp_param', `%rdx') C r8
|
|
Packit |
5c3484 |
define(`n', `%rcx') C r9
|
|
Packit |
5c3484 |
define(`u0inv', `%r8') C stack
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
define(`i', `%r11')
|
|
Packit |
5c3484 |
define(`nneg', `%r12')
|
|
Packit |
5c3484 |
define(`mp', `%r13')
|
|
Packit |
5c3484 |
define(`q0', `%rbp')
|
|
Packit |
5c3484 |
define(`vp', `%rdx')
|
|
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_redc_1)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
IFDOS(` mov 56(%rsp), %r8 ')
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
mov (up), q0 C up[0]
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
imul u0inv, q0 C first q0, for all execution paths
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov n, nneg
|
|
Packit |
5c3484 |
neg nneg
|
|
Packit |
5c3484 |
lea (mp_param,n,8), mp C mp += n
|
|
Packit |
5c3484 |
lea -16(up,n,8), up C up += n
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov R32(n), R32(%rax)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
lea 4(%rax), %r9
|
|
Packit |
5c3484 |
cmp $4, R32(n)
|
|
Packit |
5c3484 |
cmovg %r9, %rax
|
|
Packit |
5c3484 |
lea L(tab)(%rip), %r9
|
|
Packit |
5c3484 |
ifdef(`PIC',`
|
|
Packit |
5c3484 |
movslq (%r9,%rax,4), %rax
|
|
Packit |
5c3484 |
add %r9, %rax
|
|
Packit |
5c3484 |
jmp *%rax
|
|
Packit |
5c3484 |
',`
|
|
Packit |
5c3484 |
jmp *(%r9,%rax,8)
|
|
Packit |
5c3484 |
')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
JUMPTABSECT
|
|
Packit |
5c3484 |
ALIGN(8)
|
|
Packit |
5c3484 |
L(tab): JMPENT( L(0), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(1), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(2), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(3), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(0m4), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(1m4), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(2m4), L(tab))
|
|
Packit |
5c3484 |
JMPENT( L(3m4), L(tab))
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(1): CFPROT_ENDBR
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
rpm-build |
01f633 |
mov (mp_param), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 8(up), %rax
|
|
Packit |
5c3484 |
adc 16(up), %rdx
|
|
Packit |
5c3484 |
mov %rdx, (rp)
|
|
Packit |
5c3484 |
mov $0, R32(%rax)
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
jmp L(ret)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(2): CFPROT_ENDBR
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
rpm-build |
01f633 |
mov (mp_param), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r14), R32(%r14)
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mov -8(mp), %rax
|
|
Packit |
5c3484 |
mov %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add (up), %r10
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
add 8(up), %r9
|
|
Packit |
5c3484 |
adc $0, %r14
|
|
Packit |
5c3484 |
mov %r9, q0
|
|
Packit |
5c3484 |
imul u0inv, q0
|
|
Packit |
5c3484 |
mov -16(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mov -8(mp), %rax
|
|
Packit |
5c3484 |
mov %rdx, %r11
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r9, %r10
|
|
Packit |
5c3484 |
adc %rax, %r11
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
add 16(up), %r11
|
|
Packit |
5c3484 |
adc $0, %rbx
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
add %r11, %r14
|
|
Packit |
5c3484 |
adc 24(up), %rbx
|
|
Packit |
5c3484 |
mov %r14, (rp)
|
|
Packit |
5c3484 |
mov %rbx, 8(rp)
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
jmp L(ret)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(3): CFPROT_ENDBR
|
|
rpm-build |
01f633 |
mov (mp_param), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov %rax, %rbx
|
|
Packit |
5c3484 |
mov %rdx, %r10
|
|
Packit |
5c3484 |
mov -16(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r9), R32(%r9)
|
|
Packit |
5c3484 |
xor R32(%r14), R32(%r14)
|
|
Packit |
5c3484 |
add -8(up), %rbx
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov -8(mp), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add (up), %r10
|
|
Packit |
5c3484 |
mov %r10, (up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
mov %r10, q0
|
|
Packit |
5c3484 |
imul u0inv, q0
|
|
Packit |
5c3484 |
add %r9, 8(up)
|
|
Packit |
5c3484 |
adc $0, %r14
|
|
Packit |
5c3484 |
mov %r14, -8(up)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -24(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov %rax, %rbx
|
|
Packit |
5c3484 |
mov %rdx, %r10
|
|
Packit |
5c3484 |
mov -16(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r9), R32(%r9)
|
|
Packit |
5c3484 |
xor R32(%r14), R32(%r14)
|
|
Packit |
5c3484 |
add (up), %rbx
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov -8(mp), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 8(up), %r10
|
|
Packit |
5c3484 |
mov %r10, 8(up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
mov %r10, q0
|
|
Packit |
5c3484 |
imul u0inv, q0
|
|
Packit |
5c3484 |
add %r9, 16(up)
|
|
Packit |
5c3484 |
adc $0, %r14
|
|
Packit |
5c3484 |
mov %r14, (up)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -24(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov %rax, %rbx
|
|
Packit |
5c3484 |
mov %rdx, %r10
|
|
Packit |
5c3484 |
mov -16(mp), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r9), R32(%r9)
|
|
Packit |
5c3484 |
xor R32(%r14), R32(%r14)
|
|
Packit |
5c3484 |
add 8(up), %rbx
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov -8(mp), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 16(up), %r10
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
add 24(up), %r9
|
|
Packit |
5c3484 |
adc $0, %r14
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
xor R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
add -8(up), %r10
|
|
Packit |
5c3484 |
adc (up), %r9
|
|
Packit |
5c3484 |
adc 32(up), %r14
|
|
Packit |
5c3484 |
mov %r10, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
mov %r14, 16(rp)
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
jmp L(ret)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(2m4): CFPROT_ENDBR
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(lo2): mov (mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r14), R32(%r14)
|
|
Packit |
5c3484 |
xor R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mov %rax, %r10
|
|
Packit |
5c3484 |
mov 8(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov 24(up,nneg,8), %r15
|
|
Packit |
5c3484 |
mov %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 16(up,nneg,8), %r10
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov 16(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov $0, R32(%r10) C xor?
|
|
Packit |
5c3484 |
lea 2(nneg), i
|
|
Packit |
5c3484 |
add %r9, %r15
|
|
Packit |
5c3484 |
imul u0inv, %r15
|
|
Packit |
5c3484 |
jmp L(e2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(li2): add %r10, (up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov (mp,i,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
xor R32(%r10), R32(%r10)
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
L(e2): add %r9, 8(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r14
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r14, 16(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
mov 16(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %rbx, 24(up,i,8)
|
|
Packit |
5c3484 |
mov $0, R32(%r14) C zero
|
|
Packit |
5c3484 |
mov %r14, %rbx C zero
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov 24(mp,i,8), %rax
|
|
Packit |
5c3484 |
mov %r14, %r9 C zero
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add $4, i
|
|
Packit |
5c3484 |
js L(li2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(le2): add %r10, (up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %r14, %rdx
|
|
Packit |
5c3484 |
add %r9, 8(up)
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, 16(up,nneg,8) C up[0]
|
|
Packit |
5c3484 |
add $8, up
|
|
Packit |
5c3484 |
mov %r15, q0
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jnz L(lo2)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov nneg, n
|
|
Packit |
5c3484 |
sar $2, n
|
|
Packit |
5c3484 |
lea 32(up,nneg,8), up
|
|
Packit |
5c3484 |
lea (up,nneg,8), vp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -16(up), %r8
|
|
Packit |
5c3484 |
mov -8(up), %r9
|
|
Packit |
5c3484 |
add -16(vp), %r8
|
|
Packit |
5c3484 |
adc -8(vp), %r9
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
jmp L(addx)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(1m4): CFPROT_ENDBR
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(lo1): mov (mp,nneg,8), %rax
|
|
Packit |
5c3484 |
xor %r9, %r9
|
|
Packit |
5c3484 |
xor R32(%rbx), R32(%rbx)
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov %rax, %r9
|
|
Packit |
5c3484 |
mov 8(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov 24(up,nneg,8), %r15
|
|
Packit |
5c3484 |
mov %rdx, %r14
|
|
Packit |
5c3484 |
mov $0, R32(%r10) C xor?
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 16(up,nneg,8), %r9
|
|
Packit |
5c3484 |
adc %rax, %r14
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 16(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
lea 1(nneg), i
|
|
Packit |
5c3484 |
add %r14, %r15
|
|
Packit |
5c3484 |
imul u0inv, %r15
|
|
Packit |
5c3484 |
jmp L(e1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(li1): add %r10, (up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov (mp,i,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
xor R32(%r10), R32(%r10)
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r9, 8(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r14
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
L(e1): add %r14, 16(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
mov 16(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %rbx, 24(up,i,8)
|
|
Packit |
5c3484 |
mov $0, R32(%r14) C zero
|
|
Packit |
5c3484 |
mov %r14, %rbx C zero
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov 24(mp,i,8), %rax
|
|
Packit |
5c3484 |
mov %r14, %r9 C zero
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add $4, i
|
|
Packit |
5c3484 |
js L(li1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(le1): add %r10, (up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %r14, %rdx
|
|
Packit |
5c3484 |
add %r9, 8(up)
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, 16(up,nneg,8) C up[0]
|
|
Packit |
5c3484 |
add $8, up
|
|
Packit |
5c3484 |
mov %r15, q0
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jnz L(lo1)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov nneg, n
|
|
Packit |
5c3484 |
sar $2, n
|
|
Packit |
5c3484 |
lea 24(up,nneg,8), up
|
|
Packit |
5c3484 |
lea (up,nneg,8), vp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -8(up), %r8
|
|
Packit |
5c3484 |
add -8(vp), %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
jmp L(addx)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(0):
|
|
rpm-build |
01f633 |
L(0m4): CFPROT_ENDBR
|
|
rpm-build |
01f633 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(lo0): mov (mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov nneg, i
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
xor R32(%r10), R32(%r10)
|
|
Packit |
5c3484 |
mov %rax, %r14
|
|
Packit |
5c3484 |
mov %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov 24(up,nneg,8), %r15
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 16(up,nneg,8), %r14
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
add %rbx, %r15
|
|
Packit |
5c3484 |
imul u0inv, %r15
|
|
Packit |
5c3484 |
jmp L(e0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(li0): add %r10, (up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov (mp,i,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
xor R32(%r10), R32(%r10)
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r9, 8(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r14
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r14, 16(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
L(e0): mov 16(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %rbx, 24(up,i,8)
|
|
Packit |
5c3484 |
mov $0, R32(%r14) C zero
|
|
Packit |
5c3484 |
mov %r14, %rbx C zero
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov 24(mp,i,8), %rax
|
|
Packit |
5c3484 |
mov %r14, %r9 C zero
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add $4, i
|
|
Packit |
5c3484 |
js L(li0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(le0): add %r10, (up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %r14, %rdx
|
|
Packit |
5c3484 |
add %r9, 8(up)
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, 16(up,nneg,8) C up[0]
|
|
Packit |
5c3484 |
add $8, up
|
|
Packit |
5c3484 |
mov %r15, q0
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jnz L(lo0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov nneg, n
|
|
Packit |
5c3484 |
sar $2, n
|
|
Packit |
5c3484 |
clc
|
|
Packit |
5c3484 |
lea 16(up,nneg,8), up
|
|
Packit |
5c3484 |
lea (up,nneg,8), vp
|
|
Packit |
5c3484 |
jmp L(addy)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
rpm-build |
01f633 |
L(3m4): CFPROT_ENDBR
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(lo3): mov (mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
mov %rax, %rbx
|
|
Packit |
5c3484 |
mov %rdx, %r10
|
|
Packit |
5c3484 |
mov 8(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov 24(up,nneg,8), %r15
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add 16(up,nneg,8), %rbx C result is zero, might carry
|
|
Packit |
5c3484 |
mov $0, R32(%rbx) C zero
|
|
Packit |
5c3484 |
mov %rbx, %r14 C zero
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov 16(mp,nneg,8), %rax
|
|
Packit |
5c3484 |
mov %r14, %r9 C zero
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
add %r10, %r15
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
lea 3(nneg), i
|
|
Packit |
5c3484 |
imul u0inv, %r15
|
|
Packit |
5c3484 |
C jmp L(li3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(li3): add %r10, (up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
mov (mp,i,8), %rax
|
|
Packit |
5c3484 |
adc %rdx, %r14
|
|
Packit |
5c3484 |
xor R32(%r10), R32(%r10)
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r9, 8(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %r14
|
|
Packit |
5c3484 |
adc %rdx, %rbx
|
|
Packit |
5c3484 |
mov 8(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %r14, 16(up,i,8)
|
|
Packit |
5c3484 |
adc %rax, %rbx
|
|
Packit |
5c3484 |
adc %rdx, %r10
|
|
Packit |
5c3484 |
mov 16(mp,i,8), %rax
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add %rbx, 24(up,i,8)
|
|
Packit |
5c3484 |
mov $0, R32(%r14) C zero
|
|
Packit |
5c3484 |
mov %r14, %rbx C zero
|
|
Packit |
5c3484 |
adc %rax, %r10
|
|
Packit |
5c3484 |
mov 24(mp,i,8), %rax
|
|
Packit |
5c3484 |
mov %r14, %r9 C zero
|
|
Packit |
5c3484 |
adc %rdx, %r9
|
|
Packit |
5c3484 |
mul q0
|
|
Packit |
5c3484 |
add $4, i
|
|
Packit |
5c3484 |
js L(li3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(le3): add %r10, (up)
|
|
Packit |
5c3484 |
adc %rax, %r9
|
|
Packit |
5c3484 |
adc %r14, %rdx
|
|
Packit |
5c3484 |
add %r9, 8(up)
|
|
Packit |
5c3484 |
adc $0, %rdx
|
|
Packit |
5c3484 |
mov %rdx, 16(up,nneg,8) C up[0]
|
|
Packit |
5c3484 |
mov %r15, q0
|
|
Packit |
5c3484 |
lea 8(up), up
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
jnz L(lo3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C ==== Addition code ====
|
|
Packit |
5c3484 |
mov nneg, n
|
|
Packit |
5c3484 |
sar $2, n
|
|
Packit |
5c3484 |
lea 40(up,nneg,8), up
|
|
Packit |
5c3484 |
lea (up,nneg,8), vp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov -24(up), %r8
|
|
Packit |
5c3484 |
mov -16(up), %r9
|
|
Packit |
5c3484 |
mov -8(up), %r10
|
|
Packit |
5c3484 |
add -24(vp), %r8
|
|
Packit |
5c3484 |
adc -16(vp), %r9
|
|
Packit |
5c3484 |
adc -8(vp), %r10
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
lea 24(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(addx):inc n
|
|
Packit |
5c3484 |
jz L(ad3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(addy):mov (up), %r8
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
inc n
|
|
Packit |
5c3484 |
jmp L(mid)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C ALIGN(16)
|
|
Packit |
5c3484 |
L(al3): adc (vp), %r8
|
|
Packit |
5c3484 |
adc 8(vp), %r9
|
|
Packit |
5c3484 |
adc 16(vp), %r10
|
|
Packit |
5c3484 |
adc 24(vp), %r11
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
inc n
|
|
Packit |
5c3484 |
mov %r11, 24(rp)
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
L(mid): mov 16(up), %r10
|
|
Packit |
5c3484 |
mov 24(up), %r11
|
|
Packit |
5c3484 |
jnz L(al3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(ae3): adc (vp), %r8
|
|
Packit |
5c3484 |
adc 8(vp), %r9
|
|
Packit |
5c3484 |
adc 16(vp), %r10
|
|
Packit |
5c3484 |
adc 24(vp), %r11
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, 8(rp)
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
mov %r11, 24(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(ad3): mov R32(n), R32(%rax) C zero
|
|
Packit |
5c3484 |
adc R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(ret): pop %r15
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
pop %rbx
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
rpm-build |
01f633 |
CF_PROT
|