|
Packit |
5c3484 |
dnl AMD64 mpn_com optimised for CPUs with fast SSE copying and SSSE3.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2012, 2013, 2015 Free Software Foundation, Inc.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Contributed to the GNU project by Torbjorn Granlund.
|
|
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 cycles/limb cycles/limb good
|
|
Packit |
5c3484 |
C aligned unaligned best seen for cpu?
|
|
Packit |
5c3484 |
C AMD K8,K9 2.0 illop 1.0/1.0 N
|
|
Packit |
5c3484 |
C AMD K10 0.85 illop Y/N
|
|
Packit |
5c3484 |
C AMD bull 1.39 ? 1.45 Y/N
|
|
Packit |
5c3484 |
C AMD pile 0.8-1.4 0.7-1.4 Y
|
|
Packit |
5c3484 |
C AMD steam
|
|
Packit |
5c3484 |
C AMD excavator
|
|
Packit |
5c3484 |
C AMD bobcat 1.97 ? 8.17 1.5/1.5 N
|
|
Packit |
5c3484 |
C AMD jaguar 1.02 1.02 0.91/0.91 N
|
|
Packit |
5c3484 |
C Intel P4 2.26 illop Y/N
|
|
Packit |
5c3484 |
C Intel core 0.52 0.95 opt/0.74 Y
|
|
Packit |
5c3484 |
C Intel NHM 0.52 0.65 opt/opt Y
|
|
Packit |
5c3484 |
C Intel SBR 0.51 0.65 opt/opt Y
|
|
Packit |
5c3484 |
C Intel IBR 0.50 0.64 opt/0.57 Y
|
|
Packit |
5c3484 |
C Intel HWL 0.51 0.58 opt/opt Y
|
|
Packit |
5c3484 |
C Intel BWL 0.57 0.69 opt/0.65 Y
|
|
Packit |
5c3484 |
C Intel atom 1.16 1.70 opt/opt Y
|
|
Packit |
5c3484 |
C Intel SLM 1.02 1.52 N
|
|
Packit |
5c3484 |
C VIA nano 1.09 1.10 opt/opt Y
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C We use only 16-byte operations, except for unaligned top-most and bottom-most
|
|
Packit |
5c3484 |
C limbs. We use the SSSE3 palignr instruction when rp - up = 8 (mod 16). That
|
|
Packit |
5c3484 |
C instruction is better adapted to mpn_copyd's needs, we need to contort the
|
|
Packit |
5c3484 |
C code to use it here.
|
|
Packit |
5c3484 |
C
|
|
Packit |
5c3484 |
C For operands of < COM_SSE_THRESHOLD limbs, we use a plain 64-bit loop, taken
|
|
Packit |
5c3484 |
C from the x86_64 default code.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C INPUT PARAMETERS
|
|
Packit |
5c3484 |
define(`rp', `%rdi')
|
|
Packit |
5c3484 |
define(`up', `%rsi')
|
|
Packit |
5c3484 |
define(`n', `%rdx')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C There are three instructions for loading an aligned 128-bit quantity. We use
|
|
Packit |
5c3484 |
C movaps, since it has the shortest coding.
|
|
Packit |
5c3484 |
define(`movdqa', ``movaps'')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifdef(`COM_SSE_THRESHOLD',`',`define(`COM_SSE_THRESHOLD', 7)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(64)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_com)
|
|
Packit |
5c3484 |
FUNC_ENTRY(3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
cmp $COM_SSE_THRESHOLD, n
|
|
Packit |
5c3484 |
jbe L(bc)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pcmpeqb %xmm7, %xmm7 C set to 111...111
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $8, R8(rp) C is rp 16-byte aligned?
|
|
Packit |
5c3484 |
jz L(rp_aligned) C jump if rp aligned
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
lea 8(up), up
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
dec n
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(rp_aligned):
|
|
Packit |
5c3484 |
test $8, R8(up)
|
|
Packit |
5c3484 |
jnz L(uent)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifelse(eval(COM_SSE_THRESHOLD >= 8),1,
|
|
Packit |
5c3484 |
` sub $8, n',
|
|
Packit |
5c3484 |
` jmp L(am)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(atop):movdqa 0(up), %xmm0
|
|
Packit |
5c3484 |
movdqa 16(up), %xmm1
|
|
Packit |
5c3484 |
movdqa 32(up), %xmm2
|
|
Packit |
5c3484 |
movdqa 48(up), %xmm3
|
|
Packit |
5c3484 |
lea 64(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm2
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm3
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
movdqa %xmm1, 16(rp)
|
|
Packit |
5c3484 |
movdqa %xmm2, 32(rp)
|
|
Packit |
5c3484 |
movdqa %xmm3, 48(rp)
|
|
Packit |
5c3484 |
lea 64(rp), rp
|
|
Packit |
5c3484 |
L(am): sub $8, n
|
|
Packit |
5c3484 |
jnc L(atop)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $4, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa (up), %xmm0
|
|
Packit |
5c3484 |
movdqa 16(up), %xmm1
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
movdqa %xmm1, 16(rp)
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $2, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa (up), %xmm0
|
|
Packit |
5c3484 |
lea 16(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $1, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(uent):
|
|
Packit |
5c3484 |
C Code handling up - rp = 8 (mod 16)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C FIXME: The code below only handles overlap if it is close to complete, or
|
|
Packit |
5c3484 |
C quite separate: up-rp < 5 or up-up > 15 limbs
|
|
Packit |
5c3484 |
lea -40(up), %rax C 40 = 5 * GMP_LIMB_BYTES
|
|
Packit |
5c3484 |
sub rp, %rax
|
|
Packit |
5c3484 |
cmp $80, %rax C 80 = (15-5) * GMP_LIMB_BYTES
|
|
Packit |
5c3484 |
jbe L(bc) C deflect to plain loop
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sub $16, n
|
|
Packit |
5c3484 |
jc L(uend)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
movdqa 120(up), %xmm3
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
sub $16, n
|
|
Packit |
5c3484 |
jmp L(um)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(utop):movdqa 120(up), %xmm3
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, -128(rp)
|
|
Packit |
5c3484 |
sub $16, n
|
|
Packit |
5c3484 |
L(um): movdqa 104(up), %xmm2
|
|
Packit |
5c3484 |
palignr($8, %xmm2, %xmm3)
|
|
Packit |
5c3484 |
movdqa 88(up), %xmm1
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm3
|
|
Packit |
5c3484 |
movdqa %xmm3, 112(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm2)
|
|
Packit |
5c3484 |
movdqa 72(up), %xmm0
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm2
|
|
Packit |
5c3484 |
movdqa %xmm2, 96(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa 56(up), %xmm3
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
movdqa %xmm1, 80(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm0)
|
|
Packit |
5c3484 |
movdqa 40(up), %xmm2
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, 64(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm2, %xmm3)
|
|
Packit |
5c3484 |
movdqa 24(up), %xmm1
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm3
|
|
Packit |
5c3484 |
movdqa %xmm3, 48(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm2)
|
|
Packit |
5c3484 |
movdqa 8(up), %xmm0
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm2
|
|
Packit |
5c3484 |
movdqa %xmm2, 32(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa -8(up), %xmm3
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
movdqa %xmm1, 16(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm0)
|
|
Packit |
5c3484 |
lea 128(up), up
|
|
Packit |
5c3484 |
lea 128(rp), rp
|
|
Packit |
5c3484 |
jnc L(utop)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, -128(rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(uend):test $8, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa 56(up), %xmm3
|
|
Packit |
5c3484 |
movdqa 40(up), %xmm2
|
|
Packit |
5c3484 |
palignr($8, %xmm2, %xmm3)
|
|
Packit |
5c3484 |
movdqa 24(up), %xmm1
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm3
|
|
Packit |
5c3484 |
movdqa %xmm3, 48(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm2)
|
|
Packit |
5c3484 |
movdqa 8(up), %xmm0
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm2
|
|
Packit |
5c3484 |
movdqa %xmm2, 32(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa -8(up), %xmm3
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
movdqa %xmm1, 16(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm0)
|
|
Packit |
5c3484 |
lea 64(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
lea 64(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $4, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa 24(up), %xmm1
|
|
Packit |
5c3484 |
movdqa 8(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa -8(up), %xmm3
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm1
|
|
Packit |
5c3484 |
movdqa %xmm1, 16(rp)
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm0)
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $2, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa 8(up), %xmm0
|
|
Packit |
5c3484 |
movdqa -8(up), %xmm3
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm0)
|
|
Packit |
5c3484 |
lea 16(up), up
|
|
Packit |
5c3484 |
pxor %xmm7, %xmm0
|
|
Packit |
5c3484 |
movdqa %xmm0, (rp)
|
|
Packit |
5c3484 |
lea 16(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $1, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
C Basecase code. Needed for good small operands speed, not for
|
|
Packit |
5c3484 |
C correctness as the above code is currently written.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(bc): lea -8(rp), rp
|
|
Packit |
5c3484 |
sub $4, R32(n)
|
|
Packit |
5c3484 |
jc L(end)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
|
|
Packit |
5c3484 |
` ALIGN(16)')
|
|
Packit |
5c3484 |
L(top): mov (up), %r8
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
lea 32(rp), rp
|
|
Packit |
5c3484 |
mov 16(up), %r10
|
|
Packit |
5c3484 |
mov 24(up), %r11
|
|
Packit |
5c3484 |
lea 32(up), up
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
not %r10
|
|
Packit |
5c3484 |
not %r11
|
|
Packit |
5c3484 |
mov %r8, -24(rp)
|
|
Packit |
5c3484 |
mov %r9, -16(rp)
|
|
Packit |
5c3484 |
ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
|
|
Packit |
5c3484 |
` sub $4, R32(n)')
|
|
Packit |
5c3484 |
mov %r10, -8(rp)
|
|
Packit |
5c3484 |
mov %r11, (rp)
|
|
Packit |
5c3484 |
ifelse(eval(1 || COM_SSE_THRESHOLD >= 8),1,
|
|
Packit |
5c3484 |
` jnc L(top)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(end): test $1, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
mov %r8, 8(rp)
|
|
Packit |
5c3484 |
lea 8(rp), rp
|
|
Packit |
5c3484 |
lea 8(up), up
|
|
Packit |
5c3484 |
1: test $2, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
mov (up), %r8
|
|
Packit |
5c3484 |
mov 8(up), %r9
|
|
Packit |
5c3484 |
not %r8
|
|
Packit |
5c3484 |
not %r9
|
|
Packit |
5c3484 |
mov %r8, 8(rp)
|
|
Packit |
5c3484 |
mov %r9, 16(rp)
|
|
Packit |
5c3484 |
1: FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|
|
rpm-build |
c3cd4f |
CF_PROT
|