|
Packit |
5c3484 |
dnl AMD64 mpn_sublsh1_n optimised for Intel Atom.
|
|
Packit |
5c3484 |
dnl Used also for AMD bd1.
|
|
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 TODO
|
|
Packit |
5c3484 |
C * This code is slightly large at 501 bytes.
|
|
Packit |
5c3484 |
C * aorrlsh1_n.asm and this file use the same basic pattern.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C cycles/limb
|
|
Packit |
5c3484 |
C AMD K8,K9 ?
|
|
Packit |
5c3484 |
C AMD K10 ?
|
|
Packit |
5c3484 |
C AMD bd1 2.3
|
|
Packit |
5c3484 |
C AMD bobcat ?
|
|
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 (4.875 is probably possible)
|
|
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 |
define(`cy', `%r8')
|
|
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_sublsh1_n)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
xor R32(%rbp), R32(%rbp)
|
|
Packit |
5c3484 |
L(ent): mov R32(n), R32(%rax)
|
|
Packit |
5c3484 |
and $3, R32(%rax)
|
|
Packit |
5c3484 |
jz L(b0)
|
|
Packit |
5c3484 |
cmp $2, R32(%rax)
|
|
Packit |
5c3484 |
jz L(b2)
|
|
Packit |
5c3484 |
jg L(b3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b1): mov (vp), %r8
|
|
Packit |
5c3484 |
add %r8, %r8
|
|
Packit |
5c3484 |
lea 8(vp), vp
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save scy
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rbp) C restore acy
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
sbb %r8, %r15
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp) C save acy
|
|
Packit |
5c3484 |
lea 8(up), up
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
jmp L(b0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b2): mov (vp), %r8
|
|
Packit |
5c3484 |
add %r8, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
lea 16(vp), vp
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save scy
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rbp) C restore acy
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
sbb %r8, %r15
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
mov 8(up), %r15
|
|
Packit |
5c3484 |
sbb %r9, %r15
|
|
Packit |
5c3484 |
mov %r15, 8(rp)
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp) C save acy
|
|
Packit |
5c3484 |
lea 16(up), up
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
jmp L(b0)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b3): mov (vp), %r8
|
|
Packit |
5c3484 |
add %r8, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov 16(vp), %r10
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
lea 24(vp), vp
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save scy
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rbp) C restore acy
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
sbb %r8, %r15
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
mov 8(up), %r15
|
|
Packit |
5c3484 |
sbb %r9, %r15
|
|
Packit |
5c3484 |
mov %r15, 8(rp)
|
|
Packit |
5c3484 |
mov 16(up), %r15
|
|
Packit |
5c3484 |
sbb %r10, %r15
|
|
Packit |
5c3484 |
mov %r15, 16(rp)
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp) C save acy
|
|
Packit |
5c3484 |
lea 24(up), up
|
|
Packit |
5c3484 |
lea 24(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(b0): test $4, R8(n)
|
|
Packit |
5c3484 |
jz L(skp)
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax) C restore scy
|
|
Packit |
5c3484 |
mov (vp), %r8
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
mov 8(vp), %r9
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov 16(vp), %r10
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
mov 24(vp), %r11
|
|
Packit |
5c3484 |
adc %r11, %r11
|
|
Packit |
5c3484 |
lea 32(vp), vp
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax) C save scy
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rbp) C restore acy
|
|
Packit |
5c3484 |
mov (up), %r15
|
|
Packit |
5c3484 |
sbb %r8, %r15
|
|
Packit |
5c3484 |
mov %r15, (rp)
|
|
Packit |
5c3484 |
mov 8(up), %r15
|
|
Packit |
5c3484 |
sbb %r9, %r15
|
|
Packit |
5c3484 |
mov %r15, 8(rp)
|
|
Packit |
5c3484 |
mov 16(up), %r15
|
|
Packit |
5c3484 |
sbb %r10, %r15
|
|
Packit |
5c3484 |
mov %r15, 16(rp)
|
|
Packit |
5c3484 |
mov 24(up), %r15
|
|
Packit |
5c3484 |
sbb %r11, %r15
|
|
Packit |
5c3484 |
mov %r15, 24(rp)
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp) C save acy
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(skp): cmp $8, n
|
|
Packit |
5c3484 |
jl L(rtn)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
push %r12
|
|
Packit |
5c3484 |
push %r13
|
|
Packit |
5c3484 |
push %r14
|
|
Packit |
5c3484 |
push %rbx
|
|
Packit |
5c3484 |
lea -64(rp), rp
|
|
Packit |
5c3484 |
jmp L(x)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(top): mov (vp), %r8
|
|
Packit |
5c3484 |
add R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
lea 64(vp), vp
|
|
Packit |
5c3484 |
adc %r8, %r8
|
|
Packit |
5c3484 |
mov -56(vp), %r9
|
|
Packit |
5c3484 |
adc %r9, %r9
|
|
Packit |
5c3484 |
mov -48(vp), %r10
|
|
Packit |
5c3484 |
adc %r10, %r10
|
|
Packit |
5c3484 |
mov -40(vp), %r11
|
|
Packit |
5c3484 |
adc %r11, %r11
|
|
Packit |
5c3484 |
mov -32(vp), %r12
|
|
Packit |
5c3484 |
adc %r12, %r12
|
|
Packit |
5c3484 |
mov -24(vp), %r13
|
|
Packit |
5c3484 |
adc %r13, %r13
|
|
Packit |
5c3484 |
mov -16(vp), %r14
|
|
Packit |
5c3484 |
adc %r14, %r14
|
|
Packit |
5c3484 |
mov -8(vp), %r15
|
|
Packit |
5c3484 |
adc %r15, %r15
|
|
Packit |
5c3484 |
sbb R32(%rax), R32(%rax)
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rbp)
|
|
Packit |
5c3484 |
mov (up), %rbp
|
|
Packit |
5c3484 |
lea 64(rp), rp
|
|
Packit |
5c3484 |
mov 8(up), %rbx
|
|
Packit |
5c3484 |
sbb %r8, %rbp
|
|
Packit |
5c3484 |
mov 32(up), %r8
|
|
Packit |
5c3484 |
mov %rbp, (rp)
|
|
Packit |
5c3484 |
sbb %r9, %rbx
|
|
Packit |
5c3484 |
mov 16(up), %rbp
|
|
Packit |
5c3484 |
mov %rbx, 8(rp)
|
|
Packit |
5c3484 |
sbb %r10, %rbp
|
|
Packit |
5c3484 |
mov 24(up), %rbx
|
|
Packit |
5c3484 |
mov %rbp, 16(rp)
|
|
Packit |
5c3484 |
sbb %r11, %rbx
|
|
Packit |
5c3484 |
mov %rbx, 24(rp)
|
|
Packit |
5c3484 |
sbb %r12, %r8
|
|
Packit |
5c3484 |
mov 40(up), %r9
|
|
Packit |
5c3484 |
mov %r8, 32(rp)
|
|
Packit |
5c3484 |
sbb %r13, %r9
|
|
Packit |
5c3484 |
mov 48(up), %rbp
|
|
Packit |
5c3484 |
mov %r9, 40(rp)
|
|
Packit |
5c3484 |
sbb %r14, %rbp
|
|
Packit |
5c3484 |
mov 56(up), %rbx
|
|
Packit |
5c3484 |
mov %rbp, 48(rp)
|
|
Packit |
5c3484 |
sbb %r15, %rbx
|
|
Packit |
5c3484 |
lea 64(up), up
|
|
Packit |
5c3484 |
mov %rbx, 56(rp)
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp)
|
|
Packit |
5c3484 |
L(x): sub $8, n
|
|
Packit |
5c3484 |
jge L(top)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): pop %rbx
|
|
Packit |
5c3484 |
pop %r14
|
|
Packit |
5c3484 |
pop %r13
|
|
Packit |
5c3484 |
pop %r12
|
|
Packit |
5c3484 |
L(rtn):
|
|
Packit |
5c3484 |
add R32(%rbp), R32(%rax)
|
|
Packit |
5c3484 |
neg R32(%rax)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pop %r15
|
|
Packit |
5c3484 |
pop %rbp
|
|
Packit |
5c3484 |
FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
Packit |
5c3484 |
PROLOGUE(mpn_sublsh1_nc)
|
|
Packit |
5c3484 |
FUNC_ENTRY(4)
|
|
Packit |
5c3484 |
IFDOS(` mov 56(%rsp), %r8 ')
|
|
Packit |
5c3484 |
push %rbp
|
|
Packit |
5c3484 |
push %r15
|
|
Packit |
5c3484 |
neg %r8 C set CF
|
|
Packit |
5c3484 |
sbb R32(%rbp), R32(%rbp) C save acy
|
|
Packit |
5c3484 |
jmp L(ent)
|
|
Packit |
5c3484 |
EPILOGUE()
|