|
Packit |
5c3484 |
dnl AMD64 mpn_copyd optimised for CPUs with fast SSE copying and SSSE3.
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
dnl Copyright 2012, 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 0.70 0.70 Y
|
|
Packit |
5c3484 |
C AMD pile 0.68 0.68 Y
|
|
Packit |
5c3484 |
C AMD steam
|
|
Packit |
5c3484 |
C AMD excavator
|
|
Packit |
5c3484 |
C AMD bobcat 1.97 8.24 1.5/1.5 N
|
|
Packit |
5c3484 |
C AMD jaguar 0.77 0.89 0.65/opt N/Y
|
|
Packit |
5c3484 |
C Intel P4 2.26 illop Y/N
|
|
Packit |
5c3484 |
C Intel core 0.52 0.80 opt/opt Y
|
|
Packit |
5c3484 |
C Intel NHM 0.52 0.64 opt/opt Y
|
|
Packit |
5c3484 |
C Intel SBR 0.51 0.51 opt/opt Y
|
|
Packit |
5c3484 |
C Intel IBR 0.50 0.50 opt/opt Y
|
|
Packit |
5c3484 |
C Intel HWL 0.50 0.51 opt/opt Y
|
|
Packit |
5c3484 |
C Intel BWL 0.55 0.55 opt/opt Y
|
|
Packit |
5c3484 |
C Intel atom 1.16 1.66 opt/opt Y
|
|
Packit |
5c3484 |
C Intel SLM 1.02 1.04 opt/opt Y
|
|
Packit |
5c3484 |
C VIA nano 1.08 1.06 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).
|
|
Packit |
5c3484 |
C
|
|
Packit |
5c3484 |
C For operands of < COPYD_SSE_THRESHOLD limbs, we use a plain 64-bit loop,
|
|
Packit |
5c3484 |
C taken 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(`COPYD_SSE_THRESHOLD',`',`define(`COPYD_SSE_THRESHOLD', 7)')
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ASM_START()
|
|
Packit |
5c3484 |
TEXT
|
|
Packit |
5c3484 |
ALIGN(64)
|
|
Packit |
5c3484 |
PROLOGUE(mpn_copyd)
|
|
Packit |
5c3484 |
FUNC_ENTRY(3)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
lea -8(up,n,8), up
|
|
Packit |
5c3484 |
lea -8(rp,n,8), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
cmp $COPYD_SSE_THRESHOLD, n
|
|
Packit |
5c3484 |
jbe L(bc)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
test $8, R8(rp) C is rp 16-byte aligned?
|
|
Packit |
5c3484 |
jnz L(rp_aligned) C jump if rp aligned
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
mov (up), %rax C copy one limb
|
|
Packit |
5c3484 |
mov %rax, (rp)
|
|
Packit |
5c3484 |
lea -8(up), up
|
|
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 |
jz L(uent)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ifelse(eval(COPYD_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 -8(up), %xmm0
|
|
Packit |
5c3484 |
movdqa -24(up), %xmm1
|
|
Packit |
5c3484 |
movdqa -40(up), %xmm2
|
|
Packit |
5c3484 |
movdqa -56(up), %xmm3
|
|
Packit |
5c3484 |
lea -64(up), up
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
movdqa %xmm1, -24(rp)
|
|
Packit |
5c3484 |
movdqa %xmm2, -40(rp)
|
|
Packit |
5c3484 |
movdqa %xmm3, -56(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 -8(up), %xmm0
|
|
Packit |
5c3484 |
movdqa -24(up), %xmm1
|
|
Packit |
5c3484 |
lea -32(up), up
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
movdqa %xmm1, -24(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 |
lea -16(up), up
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(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 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(uent):sub $16, n
|
|
Packit |
5c3484 |
movdqa (up), %xmm0
|
|
Packit |
5c3484 |
jc L(uend)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
ALIGN(16)
|
|
Packit |
5c3484 |
L(utop):sub $16, n
|
|
Packit |
5c3484 |
movdqa -16(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
movdqa -32(up), %xmm2
|
|
Packit |
5c3484 |
palignr($8, %xmm2, %xmm1)
|
|
Packit |
5c3484 |
movdqa %xmm1, -24(rp)
|
|
Packit |
5c3484 |
movdqa -48(up), %xmm3
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm2)
|
|
Packit |
5c3484 |
movdqa %xmm2, -40(rp)
|
|
Packit |
5c3484 |
movdqa -64(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm3)
|
|
Packit |
5c3484 |
movdqa %xmm3, -56(rp)
|
|
Packit |
5c3484 |
movdqa -80(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -72(rp)
|
|
Packit |
5c3484 |
movdqa -96(up), %xmm2
|
|
Packit |
5c3484 |
palignr($8, %xmm2, %xmm1)
|
|
Packit |
5c3484 |
movdqa %xmm1, -88(rp)
|
|
Packit |
5c3484 |
movdqa -112(up), %xmm3
|
|
Packit |
5c3484 |
palignr($8, %xmm3, %xmm2)
|
|
Packit |
5c3484 |
movdqa %xmm2, -104(rp)
|
|
Packit |
5c3484 |
movdqa -128(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm3)
|
|
Packit |
5c3484 |
movdqa %xmm3, -120(rp)
|
|
Packit |
5c3484 |
lea -128(up), up
|
|
Packit |
5c3484 |
lea -128(rp), rp
|
|
Packit |
5c3484 |
jnc L(utop)
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
L(uend):test $8, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa -16(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
movdqa -32(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa %xmm1, -24(rp)
|
|
Packit |
5c3484 |
movdqa -48(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -40(rp)
|
|
Packit |
5c3484 |
movdqa -64(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa %xmm1, -56(rp)
|
|
Packit |
5c3484 |
lea -64(up), up
|
|
Packit |
5c3484 |
lea -64(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $4, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa -16(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
movdqa -32(up), %xmm0
|
|
Packit |
5c3484 |
palignr($8, %xmm0, %xmm1)
|
|
Packit |
5c3484 |
movdqa %xmm1, -24(rp)
|
|
Packit |
5c3484 |
lea -32(up), up
|
|
Packit |
5c3484 |
lea -32(rp), rp
|
|
Packit |
5c3484 |
|
|
Packit |
5c3484 |
1: test $2, R8(n)
|
|
Packit |
5c3484 |
jz 1f
|
|
Packit |
5c3484 |
movdqa -16(up), %xmm1
|
|
Packit |
5c3484 |
palignr($8, %xmm1, %xmm0)
|
|
Packit |
5c3484 |
movdqa %xmm0, -8(rp)
|
|
Packit |
5c3484 |
lea -16(up), up
|
|
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 |
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): sub $4, R32(n)
|
|
Packit |
5c3484 |
jc L(end)
|
|
Packit |
5c3484 |
|
|
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 |
mov %r8, 32(rp)
|
|
Packit |
5c3484 |
mov %r9, 24(rp)
|
|
Packit |
5c3484 |
ifelse(eval(COPYD_SSE_THRESHOLD >= 8),1,
|
|
Packit |
5c3484 |
` sub $4, R32(n)')
|
|
Packit |
5c3484 |
mov %r10, 16(rp)
|
|
Packit |
5c3484 |
mov %r11, 8(rp)
|
|
Packit |
5c3484 |
ifelse(eval(COPYD_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 |
mov %r8, (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 |
mov %r8, (rp)
|
|
Packit |
5c3484 |
mov %r9, -8(rp)
|
|
Packit |
5c3484 |
1: FUNC_EXIT()
|
|
Packit |
5c3484 |
ret
|
|
Packit |
5c3484 |
EPILOGUE()
|